⚠️ Editorial note: The open source ecosystem in China operates under a distinct institutional framework — characterized by state-led initiatives, intranet-like boundaries, and top-down governance. Readers should be aware that this context differs from the community-driven open source model common in other regions. The term “open source” as used in Chinese media may refer to practices that diverge from the conventional definition.

China Open Source Daily — 2026-11-22

🏛️ Policy & Regulation — A Root That Issues Its Own Passports: What CAICT’s Agent Identity System Publishes and What It Keeps

1. The document and the room it was said in. On 2026-11-20, the agent-internet parallel session (智能体互联网平行会议) of the 2026 China 5G+Industrial Internet Conference was held in Wuhan. Present and speaking: Liang Changdao (梁长岛), member of the party group and first-level inspector of the Hubei Provincial Communications Administration; Zheng Wanliang (郑晚亮), second-level inspector of the Hubei Provincial Department of Economy and Information Technology; and Ao Li (敖立), vice-president of the China Academy of Information and Communications Technology (CAICT, 中国信通院). The substantive presentation was by Jin Jian (金键), director of CAICT’s Institute of Industrial Internet and Internet of Things, titled “Thoughts and Practice on the Agent Identity and Identity System” (《智能体标识与身份系统思考与实践》). The article was carried by CAICT’s own account on 2026-11-20 at 23:43 and reached this run through the manual feed for 2026-11-21.

2. The diagnosis, in the report’s own words, because the wording is the argument. Jin Jian’s framing: AI is moving from single-agent intelligence to multi-agent collaboration, from closed platforms to open networks, and the consequential change is that “agents are no longer merely tools to be invoked; they begin to act in their own identity” (智能体不再只是被调用的工具,而是开始以自己的身份行动). Four questions follow and are stated as mandatory: who is acting, on whose commission, what may it do, and how is the behaviour traceable. The evaluation axis, on the report’s account, has to move from “can it do it” to “can it be trusted to do it” (从"能不能做"延伸到"能不能放心地做"). And the institutional demand is stated with unusual directness: an identity system must mark and identify agents technically, and must also establish “at the management and economic levels, which authorisations it has obtained and who bears responsibility for its behaviour and consequences” (在管理和经济层面明确其获得了哪些授权以及谁为其行为和后果负责). Technical trust plus management trust is the stated precondition for production use.

3. The architecture, as specified. CAICT proposes four principles — demand-pull, open compatibility, peer mutual recognition, interoperability (需求牵引、开放兼容、对等互认、互联互通) — and a two-tier structure. A management centre for agent identity (智能体标识与身份管理中心) forms the consensus rules: which identification methods are accepted, which credential-issuing institutions are trusted, and how widely their credentials may be used. Distributed trusted-interconnection centres (智能体可信互联中心) then provide registration, discovery, resolution and verification. The stated design aim is explicit and worth quoting: “the goal is not to bring all subjects into one centre, but to let different systems establish trust while retaining their governance autonomy” (其目标不是把所有主体纳入一个中心,而是让不同体系在保留治理自主权的基础上建立信任). Two credential species are named: an identity credential answering who it is, and a business credential further stating what it may and may not do. The summary formula is four verbs: the identifier solves discovery, identity builds trust, credentials delimit the boundary, and the result is that dispersed capabilities can collaborate.

4. What is already running, and this is the checkable part. Per the report: the agent identity service network is online and open to access; trusted-interconnection centres in Beijing, Xiamen, Chongqing and Wuhan have entered trial operation, giving a preliminary cross-regional node layout (a separate CAICT-release account reported the Wuhan centre starting trial operation on 2026-11-19, the third after Beijing and Xiamen by its count); research on architecture and regulatory governance is under way with a 《智能体互联网体系架构研究报告》 planned; domestic standards work proceeds through CCSA, international work through IETF and ITU, and CAICT has taken the lead in filing an IEEE international standard on an agent-internet reference architecture; an open-source project, OpenAgenet (OAN), is opened to industry “to provide enterprises with a technical implementation reference for understanding, validating and participating in this system”; and in April 2026 in Hong Kong CAICT co-launched a “trusted intelligent collaboration initiative” aimed at ASEAN countries, covering agent application security-capability assessment and ANP protocol interoperability validation.

5. The two sentences that carry the political economy. First: “the identity system does not grant machines autonomy detached from people, but keeps autonomous action always under human authorisation” (身份体系不是赋予机器脱离人的自主权,而是让自主行动始终处于人的授权之下) — with the companion test that a commission can be verified and revoked and behaviour traced. Second, and this is the one to keep: “global perspective, Chinese solution” (全球视角,中国方案) is CAICT’s basic philosophy for this system, justified by an analogy to passports — each party keeps independent governance the way each state keeps its own passport system, and mutual recognition proceeds under common rules — followed by the frankest line in the document: “peer mutual recognition removes collaboration obstacles; early positioning wins future initiative” (对等互认消除的是协作障碍,先发布局争取的则是未来主动权).

Institutional Analysis & Commentary:

This is the most serious institutional design document this series has read all year, and the sentence that makes it serious is the one about autonomy. It is also a document whose open-source component can be measured this morning, and the measurement is not flattering.

First, state what is genuinely new here, because the standard was already reported. This series covered GB/Z 185.2—2026, the identity-code standard, and MIIT Notice No. 42 on 2026-11-07 — the instrument that made an agent identity code a condition of a terminal’s network-access licence through the phrase “by reference to the standards”. That covered the filing regime: the code, the licence condition, the standard’s Annex B reciprocity route. What the Wuhan session adds, and what no earlier document in this series contains, is the governance architecture behind the code — the distinction between a management centre that decides whose credentials count and distributed centres that operate, the identity-credential/business-credential split, and the explicit refusal to centralise. A registry that hands out numbers is an administrative apparatus. A federation that accredits credential issuers is a constitutional design. The Wuhan document describes the second, and describes it competently.

Second, the design has three properties that a serious reader should credit, because each is a real institutional innovation and each is scarce in this ecosystem. (a) The separation of the identity credential from the business credential is the correct primitive. It corresponds, in the vocabulary of institutional economics, to distinguishing who holds the right from what the right permits — the split between entitlement and scope that every property system needs and most credentials conflate. A system that can revoke scope without revoking existence is one where a badly-behaved agent can be narrowed rather than killed, which is what real sanctions regimes look like. (b) The refusal to centralise is stated as a goal, not a concession: “not to bring all subjects into one centre.” A federation that says this on the record has given its critics a test, and this series should hold it to that test. (c) The revocation requirement is the right requirement. “Whether the commission can be verified, revoked, and its behaviour traced” is, almost verbatim, the trio that makes a delegation of authority meaningful in law. Delegation without revocation is a gift. The document names the right three.

Third, the honest and structural problem, which is where the design stops being a federation. Read the architecture again and count the parties. The management centre forms the consensus rules — and per the account, CAICT builds it. The distributed centres provide registration and verification — and the four named nodes are being stood up by CAICT and provincial partners. The root of the credential hierarchy, in the technical implementation, is Root. In the white paper that OAN publishes, the effective-authorisation formula is stated in one line: active governance state ∧ valid Root-issued credential, with Root as the issuer and verifier of protocol credentials and the chain recording only lifecycle state. That is a coherent design and it is not a federation of peers; it is a hierarchy with a documented audit trail. The white paper is candid about this and should be credited for the candour — its §27.1 lists as a material limitation that “the current architecture operates within one Root trust domain” and does not yet specify or validate how independently governed domains recognise one another’s Root records. The passport analogy in the speech is a good metaphor; the code says the passport office is one office until further notice. “Peer mutual recognition” is currently a principle with one peer.

Fourth, and now the measurement, because OAN is not a press release and can be counted. The claim in the speech is that OAN provides “a technical implementation reference” to industry. What exists, verified through the GitHub API on 2026-11-21:

  • The OpenAgenet organisation holds 20 repositories, created 2026-05-21, of which 16 are its own work and four are forks — including a fork of the W3C Decentralized Identifier Extensions registry and a fork of the AI Knowledge Representation community group’s repository.
  • Across the 16 non-fork repositories there are 488 commits authored on or after 2026-05-21. Not one of them is authored by anyone other than a single individual — the project’s initiator, whose commit identity is jlxufly@gmail.com, the same address on the white paper’s CAICT affiliation. Zero external commits in six months across the entire project.
  • The two repositories with substantial third-party history are forks of upstream W3C/DIF working documents (did-oan-extensions: 32 commits by other authors since May, of which the overwhelming majority are DID-method registration entries submitted by other projects — did:artifact, did:pki, did:sns, did:cid, did:uuid and dozens more — through the standard W3C registry pull-request flow). This is not OAN receiving contributions. This is OAN carrying someone else’s registry.
  • The organisation counts 51 to 100 total stars across all repositories; the flagship SDK, oan-sdk-ts, has 3 stars and 61 commits; the public documentation repository has 2 stars and 39 commits, last pushed 2026-11-15.
  • The SDK is published to npm as @openagenet/oan-sdk-ts: created 2026-08-01, six versions, latest 1.1.1 on 2026-10-21, and 94 downloads in the last thirty days (1,229 since August). Its LICENSE is Apache-2.0 with a carve-out that is worth quoting exactly: “OpenAgenet and OAN names, logos, slogans, official-node labels, and certification labels are not licensed under Apache-2.0.”

Fifth, the licence carve-out is the whole architecture in one paragraph, and it is a better summary than anything in the speech. The code is Apache-2.0 — genuinely open, OSI-approved, no restriction on use, modification or redistribution. The marks are not. Which means the project’s open-source status is precisely scoped: anyone may implement OAN; only the mark-holder may certify a node. That is not a defect — it is the same structure the Linux Foundation, the Open Geospatial Consortium and every serious conformance regime use, and it is the correct way to run an accreditation system. It is also the exact point at which “open” stops describing the institution and starts describing the artefact. The white paper understands this and says so under the heading “Conformance Before Branding” (§17.3), listing the conformance tests — canonicalisation, signature verification, credential proofs, Root packages, discovery authorisation, replay rejection — that a third party would have to pass. Publishing conformance tests is the single most important thing an accreditation body can do, and it is the test of whether this is a standard or a gate. Whether those tests are public and runnable by an outsider is this entry’s central open question.

Sixth, the economic sentence, which is the part a Western reader will misread. “Peer mutual recognition removes collaboration obstacles; early positioning wins future initiative” is not cynicism and it is not a policy slogan. It is an accurate statement of standard-competition economics: whoever runs the accredited root of a federation collects the coordination rents that accrue at the root, and the window to establish that root closes when a competing regime achieves scale. Every standards body in history has known this — the IETF’s objection to ISO’s OSI was about who would sit at the root of the naming hierarchy, not about elegance. The uncomfortable observation for the ecosystem this series covers is different: here the party that would sit at the root is a state research institute, and the open-source project that would carry the specification has attracted no outside contributor in six months and ninety-four downloads a month. For comparison, this series documented on 2026-11-21 that the LoongArch community’s flagship repository drew one outside pull request in its first forty-three days and this briefing called that figure thin. OAN has had one hundred and eighty-four days and zero. A standard can be imposed through a licence condition; it cannot be made a commons by publication alone.

One perspective, not a verdict. Each question below is answerable from a public record within the quarter:

(i) Are OAN’s conformance tests public and runnable by a party that is not a CAICT partner? §17.3 describes the test surfaces. A test suite a third party can execute is a standard; a test suite only the root can run is an examination. The difference is one repository’s visibility setting.

(ii) Will the governance layer’s root be operated by one entity permanently, or does the white paper’s §27.1 limitation have a roadmap with a date? Cross-domain recognition among independently governed roots is the difference between a national system with a reciprocity annex and a genuine federation. The white paper lists the gap; the roadmap is the answer.

(iii) Does the “trusted credential issuer” list publish its criteria and its members? The management centre’s most consequential power is deciding whose credentials are trusted in what scope. That is a decision about market access taken by an unlisted body, and it is the one place in this architecture where the federation’s autonomy claim could quietly fail.

(iv) Will OAN ever accept a commit from someone outside the initiating organisation? The number is currently zero out of 488. A first outside pull request merged by a maintainer who is not the author would be the smallest possible evidence that this is infrastructure other people build on rather than a specification one person publishes.

Sources: 中国信通院 CAICT via WeChat — 中国信通院金键:智能体标识与身份系统是AI关键基础设施 (2026-11-20 23:43; the 5G+Industrial Internet Conference agent-internet parallel session in Wuhan on 2026-11-20; Liang Changdao, Zheng Wanliang and Ao Li in attendance; Jin Jian’s keynote; the shift from single-agent to multi-agent and from closed platforms to open networks; agents acting in their own identity; the four mandatory questions of who acts, on whose commission, what it may do and how behaviour is traced; the move from “can it do it” to “can it be trusted to do it”; the management-and-economic-layer authorisation requirement; the four principles of demand-pull, open compatibility, peer mutual recognition and interoperability; the management centre deciding accepted identification methods and trusted credential issuers and credential scope; the distributed trusted-interconnection centres; the refusal to centralise and the retention of governance autonomy; the identity credential and business credential distinction; the online service network with Beijing, Xiamen, Chongqing and Wuhan centres in trial operation; the architecture research report, CCSA domestic standards, IETF and ITU participation and the filed IEEE reference-architecture standard; the OpenAgenet (OAN) open-source project; the April 2026 Hong Kong initiative with ASEAN security assessment and ANP validation; the “human authorisation” formulation with verify, revoke and trace; the passport analogy; and the “early positioning wins future initiative” line) · 证券时报 / 人民财讯 — 中国信通院智能体标识与身份网络面向社会开放接入 (2026-10-09; the network’s trial operation and open access, with registration, identity verification, addressing resolution, capability discovery and collaborative verification services, initially through the Beijing and Xiamen centres) · OpenAgenet / OAN White Paper — arXiv:2606.03161v8 (Jinliang Xu, China Academy of Information and Communications Technology; submitted 2026-06-02, v8 revised 2026-11-12; the identity, governance, discovery and freshness gaps; did:oan resource identity and governance-backed admission; Root, Registrar, Discovery and CDN roles; the effective-authorisation formula of active governance state AND a valid Root-issued credential; infrastructure authorization and the on-chain governance layer for lifecycle events; auditability requirements; “Conformance Before Branding” and the conformance test surfaces; §26 on avoiding lock-in and private roots; and §27.1’s stated limitation of a single Root trust domain) · GitHub — OpenAgenet organisation and repository metadata (verified 2026-11-21 through the GitHub API: 20 repositories created from 2026-05-21, 16 non-forks, 488 commits from non-fork repositories dated on or after 2026-05-21 with zero authored by anyone other than the initiating account, 51–100 total stars, oan-sdk-ts at 3 stars and 61 commits, oan-public-docs at 2 stars and 39 commits last pushed 2026-11-15, and the Apache-2.0 licence with a mark-and-certification-label carve-out) · npm registry — @openagenet/oan-sdk-ts (published 2026-08-01, six versions, latest 1.1.1 on 2026-10-21, 94 downloads in the thirty days to 2026-11-20 and 1,229 since publication) · This series’ 2026-11-07 briefing — the identity code becomes a condition of market access: MIIT Notice No. 42 and GB/Z 185.2—2026


🏗️ Institutional Change — Seventeen Point Two Seven Per Cent and No Votes: What DeepSeek’s Closed Second Round Says About the Ownership of Open Weights

1. The event, with the date that matters. On 2026-11-17, DeepSeek’s business-registration change was finalised. The change records that the new investors in DeepSeek’s second external round took a stake of approximately 17.27 per cent in the company, and that DeepSeek’s registered capital rose from about CNY 16.45 million to CNY 19.92 million, an increase of 21.1 per cent. Caixin reported on 2026-11-20 that the round closed at approximately CNY 100 billion (USD 15 billion) and that DeepSeek is preparing an IPO on Shanghai’s STAR Market. The share-platform detail, from Cailian Press on 2026-11-19: the new external shareholders are held through Hangzhou Chengyan Enterprise Management Consulting Partnership (杭州程砚企业管理咨询合伙企业), whose registered capital contribution rose from CNY 6,000 to about CNY 3.44 billion, with its partner count rising to 32 — the 31 new partners being the round’s external investors. At a post-money valuation of CNY 615.5 billion, the 17.27 per cent corresponds to about CNY 106.3 billion, roughly matching the CNY 105.5 billion round.

2. The structure, which is the substantive news rather than the number. The instrument is a limited partnership platform sitting between the investors and the company. Its properties, per the reporting: external capital holds economic rights through a limited-partnership structure, does not enjoy company voting rights, and is subject to a five-year lock-up (the arrangement described for the first round and carried into the second). The founder Liang Wenfeng (梁文锋) remains the controlling shareholder. The first round — closed at the end of June 2026, CNY 51 billion raised, post-money valuation near CNY 400 billion, with Liang himself contributing roughly CNY 20 billion as the largest single subscriber and the National AI Industry Investment Fund taking a small stake — established the template: money in, votes not.

3. The investors, because their composition is the institutional content. Per Cailian Press’s registry analysis, at least nine securities firms participated: CITIC Securities, CICC, Guoyuan, Caitong, Orient, Zheshang, Shenwan Hongyuan, Huafu and CSC. Together they hold 9.02 per cent of the Hangzhou Chengyan platform, equating to 1.558 per cent of DeepSeek and about CNY 9.51 billion of the round. CITIC Securities alone accounts for 84.8 per cent of the brokerage total — about CNY 8.07 billion through its wholly-owned alternative-investment subsidiary, equivalent to 62 per cent of that subsidiary’s registered capital. CATL’s system holds 9.5653 per cent of the platform through three funds, tying with Tencent as the largest external holders. China Mobile Investment holds a position comparable to CITIC’s. The remainder came through funds managed by outside GPs — Chengdu Shixi, Monolith, CPE Yuangfeng, Hefei state capital platforms and others.

4. The conflicts that the structure does not resolve, stated as the reporting states them. CITIC Securities is simultaneously a shareholder of DeepSeek and its chosen listing sponsor for the STAR Market. Cailian Press names this the “investment banking plus investing” model and observes that CITIC stands to collect investment returns, underwriting and sponsorship fees, and the follow-on allocation from the same transaction. Separately, the reporting notes that the round’s investors face a five-year lock-up with no voting rights, that pre-IPO holdings are carried at cost or equity method and do not track secondary prices, and that the realisation of any return depends entirely on the IPO window. On the management side: Yan Wentao (严文韬), formerly a partner at GL Ventures (高瓴创投) — which is not itself a DeepSeek investor and where he never led a DeepSeek investment — joined as the company’s first-ever CFO in September 2026, a role DeepSeek had publicly advertised since February 2025 and still had not filled in its June 2026 recruitment notice. His mandate is conventionally read as IPO preparation.

Institutional Analysis & Commentary:

The most consequential open-weight model family in the world has just been capitalised through an instrument that deliberately separates ownership from control — and the separation is not a side effect of the paperwork. It is the point.

First, name the instrument correctly, because the label decides the analysis. A limited-partnership holding platform with economic rights, no voting rights, and a five-year lock-up is not a share issue. It is a revenue-and-residual claim with a fixed term — closer in economic substance to a convertible instrument with a toll on the exit than to equity in a governed company. In Williamson’s terms, this is a governance structure chosen to solve one specific hazard: the investor’s need for liquidity and the founder’s need for autonomy are incompatible, and the platform lets both be satisfied by giving each party a different bundle of rights over the same assets. The founder keeps control because control is the thing he cannot buy back later; the investors get a defined share of the upside because that is the thing they can wait five years for. This is a legitimate and sophisticated solution to a real problem, and it is increasingly the standard form for Chinese AI companies that intend to list while their founders intend to keep deciding. Reading it as a corporate-governance failure would be a mistake. Reading it as a choice about whose preferences bind is exact.

Second, and this is what changes the meaning of everything this series has written about open weights: it answers who owns the artefact, and it answers it against the commons. DeepSeek’s weights — V4, the million-context variants, the whole line that this series documented on 2026-11-20 as carrying 15 per cent of enterprise open-source spend worldwide with no commercial restrictions — were the base of the domestic market and, per Menlo Ventures’ data, one of four Chinese vendors holding 89 per cent of that spend. The licence imposes nothing. The weights are free. And the entity that will decide the next release’s licence, the next model’s publication, and whether any of it stays open after listing is a company whose capital structure now gives thirty-one outside investors a claim on residual value and not one of them a vote on the decision. That is not hypocrisy; the open licence was always a choice the company made and can stop making. It is the plainest possible illustration of the fact this series keeps arriving at: an open-weight release transfers an artefact, not a governance right, and the governance stays wherever the share register is. Every downstream adopter who built a product on a DeepSeek model, every enterprise that chose it over a paid alternative on the strength of the licence, holds a dependency whose future licensing is decided by a board they cannot influence and a founder they cannot replace.

Third, the CITIC dual role deserves a careful sentence rather than an accusation, because the pattern is older and wider than this deal. A securities firm that invests and sponsors is doing what universal banks have done for a century, under disclosure. The institutional question is not whether the conflict exists — it exists, and it is disclosed — but whether the party that benefits from a high listing valuation is the same party that certifies the company’s financial statements to the exchange. The sponsor’s due diligence is the buyer’s assurance. When the sponsor is also a shareholder with a five-year lock-up, its incentive to overstate is reduced, since it cannot sell into the pop. That is a real mitigation and it should be credited. But note the direction of the residual: the sponsor’s return is maximised by a successful listing and a stable aftermarket, which means its interest in prompt and full disclosure conflicts with its interest in a smooth process — and in a STAR Market listing where the issuer’s product is a model whose revenue is API calls, the disclosure questions are unusually hard. This briefing has no evidence of misconduct and asserts none. It records that the same institution is on both sides of the most consequential disclosure in the Chinese AI market this year, and that the arrangement was designed that way.

Fourth, connect it to yesterday’s finding, because the two documents are the same document read at different altitudes. On 2026-11-21 this series measured the LoongArch community at 22 founding members and six commits, and argued that a founding roster measures intent while a commit log measures capacity. Today’s story is the capital-market edition of the same measurement. DeepSeek’s round: 31 external investors, CNY 105.5 billion, 17.27 per cent, zero votes. Count the inputs and the intake is enormous. Ask what the investors can decide and the answer is nothing. Two of the largest institutional arrangements in Chinese technology this week are instruments that take in a great deal and confer no decision rights — one absorbing members into a community charter, the other absorbing capital into a partnership platform. Neither is illegitimate. Both share a structural signature: the perimeter expands, the centre does not move. That is the pattern this series should carry forward as its working hypothesis about Chinese institutional forms of this period: intake is generous, authority is concentrated, and the two are compatible because the thing being distributed is participation rather than power.

Fifth, the honest caveat, and it cuts against the moralising reading. Unanimity was never on offer, and pluralism is not obviously better here. DeepSeek’s competitive advantage has been decisiveness — the release cadence, the licence choices, the refusal to enter the API price war on anyone else’s terms, the free-tier limits this series examined on 2026-11-17. A company that has to convene thirty-one limited partners before releasing weights would not have released them. The absence of votes is also the absence of a veto. Founders who raise capital without ceding control are making a defensible bet that speed beats consensus in a market where the frontier moves quarterly — and the empirical record of founder-controlled AI labs, in China and elsewhere, is not uniformly worse than that of board-governed ones. What is wrong is not the arrangement. What is wrong is claiming the arrangement renders the commons durable. It renders the company fast.

Sixth, what to watch, all of it observable. Does the 2027 prospectus disclose the K3-style revenue thresholds and licence terms as related-party or risk matters — and does it describe the licence as a policy the company reserves the right to change? Does the National AI Industry Investment Fund’s stake in a company with a permissive licence create any governance expectation about continued openness, or is it purely financial? And: what happens to the licence terms of the model released after the listing — a permissive release, as before, or a threshold like the ones Alibaba and MiniMax adopted? The first post-IPO licence is the document that tells the whole world what an open-weight commitment was worth.

Checkable questions.

(i) Does the Hangzhou Chengyan platform’s partnership agreement contain any information rights, board observer seats, or consent matters attached to the economic stake? “No voting rights” is stated; the full bundle of non-voting rights — information, consultation, consent on a sale — is not, and it is where real influence lives.

(ii) What is CITIC’s total remuneration across the three streams — investment return, sponsorship fee, and follow-on allocation? One number would make the conflict legible. It will not appear until the prospectus.

(iii) Do any of the nine brokerage investors intend to distribute their stakes to fund investors before the lock-up expires, and what does that imply for the platform’s stability? A limited-partnership layer can be unwound by assignment where shares cannot.

Sources: Caixin Global — DeepSeek Closes 100 Billion Yuan Funding Round (2026-11-20; the approximately CNY 100 billion / USD 15 billion second round; business-registration changes finalised on 2026-11-17; the new-investor partnership taking roughly 17.3 per cent; the first round’s CNY 50 billion at a CNY 350 billion post-investment valuation with registration completed in July; the STAR Market IPO objective) · 财联社 / 21世纪经济报道 — 9家券商参投DeepSeek,对应出资95亿,中信81亿占大头 (2026-11-19 and 2026-11-19 22:21; registered capital from about CNY 16.45 million to CNY 19.9187 million; the Hangzhou Chengyan platform’s contribution rising from CNY 6,000 to about CNY 3.44 billion with 32 partners and 31 new external investors; the 17.27 per cent stake and the CNY 6,155 billion post-money valuation against a CNY 105.5 billion round; the nine securities firms and their 9.02 per cent platform share, 1.558 per cent DeepSeek share and CNY 9.51 billion contribution; CITIC Securities at 84.8 per cent of the brokerage total, about CNY 8.07 billion, and 62 per cent of its subsidiary’s registered capital; China Mobile Investment’s comparable position; CATL at 9.5653 per cent tying with Tencent; the GP routes through Chengdu Shixi, Monolith, CPE and others; CITIC as both investor and listing sponsor and the three revenue streams; and the five-year lock-up, absence of voting rights and the cost-method and equity-method accounting with liquidity discounts) · 21世纪经济报道 — 高瓴创投原合伙人严文韬正式入职DeepSeek (2026-09-23 and 2026-09-24; Yan Wentao’s appointment as CFO and his GL Ventures background; the first round’s investor list including Liang Wenfeng, Guozhi Investment, Tencent, NetEase, JD, the CATL system, Zhengxin Gu, Shixiang Technology, Monolith and IDG Capital; about CNY 51 billion raised at a post-money valuation near CNY 400 billion with Liang contributing about CNY 20 billion; the limited-partnership structure with economic rights only, no voting rights and a five-year lock-up; the National AI Industry Investment Fund taking 0.28 per cent; and CITIC Securities hired on 2026-09-09 to prepare a STAR Market listing) · Reuters — China’s DeepSeek hires dealmaker as CFO ahead of possible IPO (2026-09-14; Yan Wentao’s appointment, CITIC Securities’ engagement, and the transition from a research lab bankrolled by its founder’s hedge fund to a conventional corporate structure) · This series’ 2026-11-20 briefing — twenty-two times the volume and forty-five times less the money


1. The document. The LICENSE file of github.com/OpenAgenet/oan-sdk-ts, read on 2026-11-21. It is six lines long and it performs two operations in sequence. First: SPDX-License-Identifier: Apache-2.0, with a pointer to the canonical text and the statement that “this repository is licensed under the Apache License, Version 2.0.” Second, immediately after, with no transition: "OpenAgenet and OAN names, logos, slogans, official-node labels, and certification labels are not licensed under Apache-2.0." Across the organisation, the licence situation is mixed and mostly deliberate: oan-bulletin-contracts, oan-community-skill, oan-identity-gateway, oan-identity-runtime, oan-identity-sidecar, oan-did-resolver and universal-resolver-for-did-oan carry full Apache-2.0 text; oan-sdk-ts, oan-public-docs, oan-trust-indexer, oan-root-nodes, oan-registrar-nodes, oan-discovery-nodes and did-oan-extensions return NOASSERTION to GitHub’s licence detector because their Apache header is accompanied by the carve-out; and oan-identity-lifecycle-evaluation, oan-demos and two experiments repositories carry no licence file at all.

2. Why this is a licensing story and not a footnote. The Apache-2.0 grant is real and generous: it permits use, modification, distribution and sublicensing, including in proprietary products, with a patent grant and a patent-retaliation clause. It does not, and cannot, extend to trademarks — Apache-2.0 §6 says so explicitly. What the carve-out does is take a right that was already excluded by the licence and put it in the same file, one line below the grant, so that every reader encounters the two together. That is not a trick. It is good drafting for a standards body, because the single most common failure mode in open specification work is a downstream adopter who reads “Apache-2.0” and assumes they may call their implementation certified. The file forecloses that misreading in the place where the misreading would occur.

3. What the white paper says should accompany the mark. The OAN paper devotes §17.3 to “Conformance Before Branding”, listing what conformance tests must cover: object canonicalisation, signature verification, credential proofs, Root packages, discovery authorisation, current-version handling and replay rejection, plus DID Document proofs, identity generation and import, controller authorisation, external identifier handling and public-key discovery for application authentication. §12.5 specifies the audit path an operator must be able to reconstruct: who submitted an identity, which Registrar issued the credential, which Root decision accepted it, which package was published, which Discovery node indexed it, and why a request was accepted or rejected. Those are the right tests. They are also, conspicuously, tests that only a party holding the Root’s keys and package formats can run end to end.

Institutional Analysis & Commentary:

The open-source strategy of the most important identity infrastructure initiative in China this year is a permissive licence on the code with the certification right retained — which is the correct strategy, the standard strategy, and, on the evidence, a strategy that is not working.

First, be exact about what the arrangement is, because it is routinely misread in both directions. A reader inclined to suspicion will see “open core” or “fake open source.” That reading is wrong: the Apache-2.0 grant here is unqualified in scope — no field-of-use restriction, no revenue trigger, no commercial annexe, unlike the licence structure this series examined on 2026-11-20 in which a Chinese model vendor conditioned model-as-a-service rights on a USD 20 million revenue threshold. Anyone may take this SDK, fork it, ship it in a product and never speak to CAICT again. A reader inclined to generosity will say the mark reservation is irrelevant because the code is what matters. That reading is also wrong: in an accreditation architecture, the mark is the product, and §12 tells you why — the whole point of OAN is that a node presenting a credential has been verified by a Root. Without the mark, a conforming implementation is indistinguishable from a non-conforming one, and the trust layer provides no trust. Both halves are load-bearing, and the licence states them both.

Second, and this is the finding the numbers force: the strategy’s success condition is adoption, and adoption is not happening at a scale a licence can fix. Forty-eighty-eight commits across sixteen repositories in six months, all from one email address. Ninety-four npm downloads in the last thirty days; 1,229 in total since August. Three stars on the SDK. Two stars on the public documentation. For a project whose stated purpose is to provide industry with a technical implementation reference for an infrastructure the state is standing up, these are not the numbers of an emerging standard. They are the numbers of a well-documented internal implementation that has been published. And that matters beyond vanity metrics for one structural reason: a conformance regime’s authority rests on independent implementations existing. If nobody outside the root builds a node, the “conformance test” verifies nothing — there is nothing to conform. The white paper’s §26 imagines a plural ecosystem: “a public trust domain may operate one Root with multiple independent Registrars and Discovery nodes; an enterprise may operate a private Root; a research testbed may allow experimental roots; a standards community may use OAN object models and conformance tests without mandating a single public operator.” That paragraph is the best thing in the document, and at present every clause of it describes an organisation that has not arrived.

Third, connect the licence to the architecture and a sharper point appears: this is copyleft’s mirror image, and its weakness is the same weakness in reverse. Copyleft — GPL, and the Mulan-series’ variants this series has tracked in Chinese foundation work — retains control by extending obligations downstream: you may use it, but what you distribute must stay open, and the obligation travels with the code. The OAN arrangement retains control by retaining nothing downstream and everything at the root: you may use it entirely freely, and the only thing you cannot do is claim the credential the architecture depends on. Copyleft’s enforcement problem is that obligations are hard to detect and expensive to litigate. This arrangement’s enforcement problem is worse and different: the right retained is worthless unless people want it. A certification nobody applies for is not a monopoly rent; it is an unstaffed queue. The licence chose the correct instrument for a standards body and inadvertently revealed that the standard has no aspirants.

Fourth, the governance consequence of the licence-mix, which is an actual practical hazard and not a rhetorical one. Four repositories carry no licence file at all (oan-demos, oan-identity-lifecycle-evaluation, and the two experiments-* repositories), and seven more return NOASSERTION because of the carve-out. For a normal project this would be housekeeping. For a project whose purpose is to be cited by enterprises as a technical reference for a regulated identity system, the consequence is concrete: a corporate counsel conducting due diligence on OAN will reject the unlicensed repositories outright — default copyright means no rights are granted at all — and will escalate the seven NOASSERTION repositories, because GitHub’s detector cannot certify them and the lawyer will not read a six-line file to guess. The white paper’s own §18 implementation guidance tells enterprises to “start with the trust path,” which is exactly the code path that lives in the NOASSERTION repositories. The most sophisticated institutional design document this series has read this year is accompanied by a licensing inventory that would fail a routine procurement review, and the fix is an afternoon’s work: a full Apache-2.0 LICENSE in every repository, the mark carve-out moved into a separate TRADEMARK.md, and an SPDX expression in each package manifest. This is the cheapest available improvement to a project’s institutional credibility, and it is the one nobody in the ecosystem ever does.

Fifth, the genuinely interesting question this raises about Chinese open-source strategy, which is where the entry earns its place. This series has documented, at length, the ecosystem’s preference for instruments with a gate: procurement lists, certification regimes, licence conditions, contribution-oriented talent evaluation that has no verifier. Here is a project built by a state research institute, on the most internationally legible legal footing available — an OSI-approved permissive licence with a standards-conforming trademark policy — carrying the governance design that the national standard’s Annex B anticipated when it offered a reciprocity route for foreign jurisdictions. By the standards this series has been applying all year, this is the most internationally compatible artefact the Chinese institutional system has produced. And it has zero outside contributors and ninety-four downloads a month. The lesson is not that the Chinese system cannot produce open institutions. It is that producing them is not the binding constraint; attracting anyone to them is — and the reason nobody came is not the licence, the mark, or the governance text, all of which are better than average. It is that publication is not distribution, and a paper on arXiv and a repository on GitHub do not by themselves create a community. That is a problem the Chinese ecosystem and the Western one share, and it is the same problem this series identified in the LoongArch community thirty-six hours ago, stated in a different key.

Checkable questions.

(i) Is there a TRADEMARK.md or a published trademark policy stating what “certified” requires and who may use the mark? The licence says the marks are reserved; a reserved right without a stated route to obtain it is a permanent exclusion.

(ii) Will the four unlicensed repositories receive licences, and will the seven NOASSERTION repositories be normalised to a machine-readable SPDX form? This is the difference between a project an enterprise’s counsel can clear in an hour and one that stalls a procurement cycle.

(iii) Has any implementer outside the initiating organisation passed the §17.3 conformance tests? One passing third-party implementation converts an accreditation scheme into a standard.

Sources: GitHub — OpenAgenet/oan-sdk-ts LICENSE (the Apache-2.0 SPDX identifier and grant, followed by the exclusion of OpenAgenet and OAN names, logos, slogans, official-node labels and certification labels from the Apache-2.0 grant) · GitHub API — OpenAgenet organisation licence inventory (verified 2026-11-21: Apache-2.0 text in oan-bulletin-contracts, oan-community-skill, oan-identity-gateway, oan-identity-runtime, oan-identity-sidecar, oan-did-resolver and universal-resolver-for-did-oan; NOASSERTION in oan-sdk-ts, oan-public-docs, oan-trust-indexer, oan-root-nodes, oan-registrar-nodes, oan-discovery-nodes and did-oan-extensions; no licence in oan-demos, oan-identity-lifecycle-evaluation and the two experiments-* repositories) · OpenAgenet / OAN White Paper — arXiv:2606.03161v8 (§12.5 auditability, §17.3 conformance-before-branding test surfaces, §18.1 start-with-the-trust-path implementation guidance, and §26’s plural-ecosystem model of public, enterprise, testbed and standards-community roots) · This series’ 2026-11-20 briefing — a toll booth in a licence: the K3 revenue threshold and Moonshot’s metering problem


🔍 WeChat Monitor

Two items were present in the manual feed for 2026-11-21 (china-daily-english-input, raw file raw/2026-11-21.md, repository pulled and found up to date at the time of this run, raw/ last modified 2026-11-21 18:07).

  • 中国信通院 (CAICT official account), 中国信通院金键:智能体标识与身份系统是AI关键基础设施 — 2,179 characters of body text, published 2026-11-20 23:43. Extracted in full and used as the primary source for the Policy & Regulation entry: the Wuhan conference session and its attendee list, the four mandatory questions, the four principles, the management-centre/distributed-centre split, the identity-credential and business-credential distinction, the four trial-operation cities, the IEEE filing, the OpenAgenet project, the Hong Kong/ASEAN initiative, the “under human authorisation” formulation and the “early positioning” line. Cross-checked against CAICT’s own October release on the network’s open access and against the OAN white paper.
  • 证券时报 (Securities Times) via 经济观察报, 武汉智能体可信互联中心启动试运行 — carried by QQ News at 2026-11-20 16:52, describing the Wuhan centre’s trial operation beginning on 2026-11-19 as the third such centre after Beijing and Xiamen, with CAICT chief engineer Ao Li and Wuhan municipal officials at the launch. Note the one-city discrepancy worth recording: this report counts Wuhan as the third centre (Beijing, Xiamen, Wuhan), while the CAICT account lists four cities with Chongqing and places Wuhan’s trial operation at the conference. The discrepancy is small and probably reflects different start dates; it is recorded rather than resolved.

The feed’s structured slots for the monitored accounts (OpenAtom, CCF’s open-source development committee, Huawei Open Source, Mulan, COPU, Tiangong Kaiwu, Ming Says, BAAI FlagOpen, KAIYUANSHE, Silicon Time, AtomGit, the Linux Kernel Developers Conference, GOSIM/CSDN, the Huang Daniang Tea-Swotting House and Yetian’s Book) were all empty for the day. No item was used from the previous three days’ feeds.

Dedup against the previous three briefings, stated precisely. The 2026-11-21 entry covered the OpenAtom Zhengzhou conference, the LoongArch community’s six commits, the OpenAtom statutory filings, and the secure-and-reliable assessment’s first mass expiry; none of those recrosses below. The 2026-11-20 entry covered Moonshot’s K3 revenue share, Shanghai’s Model Ark import and the Menlo Ventures report; none recrosses. The 2026-11-19 entry covered Bank of Communications’ CAICT governance assessment, Moonshot’s repapering and Zhipu’s valuation — today’s CAICT material is a different instrument at the same institution (an infrastructure and accreditation architecture rather than an enterprise governance assessment), built on a different document with a different date, and the BoCom grade is not re-reported. Agent identity has been covered before, and the relationship must be stated rather than glossed: the 2026-11-07 entry analysed MIIT Notice No. 42 and GB/Z 185.2—2026 — the filing regime, the network-access licence condition, the identity-code standard’s Annex B reciprocity route. Today’s entry is the architecture behind that standard: the management-centre/distributed-centre split, the credential species, and the OAN implementation whose commit log is measured here for the first time. OpenAgenet, OAN, did:oan, the management centre, the four trial cities, the IEEE reference-architecture filing and the Hong Kong/ASEAN initiative have not appeared in this series before — grep across the November entries and ~/.hermes/cron/output/bf436b4758a0/ returned no prior mention. DeepSeek’s second round was referenced in passing in the 2026-11-19 entry (as a cited figure of CNY 105.5 billion raised at a CNY 615.5 billion post-money valuation) and its free-tier limits were analysed on 2026-11-17; the share register — the Hangzhou Chengyan platform, the 31 external partners, the 17.27 per cent, the five-year lock-up without voting rights, the nine brokerage participants and CITIC’s dual role — has not appeared before, and the authoritative Caixin confirmation of closure on 2026-11-17 is new. The licence analysis in Legal & Licensing is entirely new and is built on a file read this morning.

🔍 Commentary — Two Perimeters, One Centre: What This Week’s Instruments Distribute and What They Keep

First: today’s two documents are the same instrument applied to two different goods, and neither of them is a commons. OAN distributes a specification — Apache-2.0 code, a published white paper, a conformance regime, an open invitation to build nodes — and retains the root of the credential hierarchy and the certification mark. DeepSeek’s round distributes capital access — CNY 105.5 billion of it, across thirty-one investors — and retains every voting right. In both cases the perimeter widens impressively and the centre does not move. In both cases that is defensible on its own terms, and in both cases the defensibility rests on the same argument: the thing being distributed is participation, and the thing being retained is the decision. This series has spent six weeks cataloguing Chinese instruments that grant entry and withhold authority. Today’s two documents are the clearest specimens yet, because neither of them is a government document — one is a research institute’s open-source project, the other is a founder’s term sheet — and both reached the same structure independently. That is evidence about the environment, not about the actors.

Second, and this is where the instinct to moralise should be resisted hardest: the retention of the centre is often the correct engineering choice. A federation whose root is contested cannot revoke a compromised node. A founder who must clear a release with thirty-one limited partners cannot ship on the cadence that made the company valuable. The question is never “should authority be concentrated” — it is “what is the cheapest mechanism that lets authority be checked.” And there are two answers, both visible in today’s documents, both cheap. For OAN: publish the conformance tests and the trademark policy. A root that anyone may test against is a root that can be audited; a root that anyone may apply to certify is a root that can be contested. Neither requires giving up a key or a mark. For DeepSeek: disclose the licence policy as a policy. If the permissive licence is a commitment, say what would change it, and let the market price that statement. Neither of these is a radical demand. Both of them are the sort of thing that a mature institution does automatically and a young one forgets.

Third, the week’s most interesting number is not any of the large ones. Not CNY 105.5 billion, not 17.27 per cent, not 488 commits, not 94 downloads. It is one: the number of outside pull requests in the LoongArch community’s flagship repository in its first forty-three days, which this series reported thirty-six hours ago and which is still the highest external-contribution figure of any newly chartered Chinese open community examined this month. OAN’s figure is zero over 184 days. The compare matters because both projects were launched with substantial institutional backing, both publish citable governance documents, both carry licences an outside lawyer can clear, and neither has attracted a single sustained outside contributor. The pattern is not about the licences and not about the governance text. It is about the absence of anything a stranger would want to build on — a problem, a dataset, a workload, a paying user. A community chartered before anyone has a reason to join it is a room with the chairs arranged. That is the finding this week’s evidence supports, and it applies to Zhengzhou and to Wuhan equally.

Fourth, the connective tissue with everything this series has argued since October, and it is now measurable in both directions. The thesis has been: open artefacts transfer, governance does not. Today it acquires a licence-level and a cap-table-level proof. OAN transfers the code and retains the credential — the artefact is open, the trust is not. DeepSeek transfers the weights and retains the board — the model is free, the future licensing is not. Neither party is doing anything wrong, and both parties would say the arrangement is open — accurately, by the definition each is using. What the Chinese ecosystem and the Western one share, and what neither has yet built, is an institution that makes the retained part contestable: a foundation with a published membership, a notice period and an appeal; a conformance regime anyone can run; a licence policy with stated amendment conditions; a cap table where the people who depend on the artefact have some say in whether it stays open. This week produced two excellent answers to “how do we build this fast.” Neither produced an answer to “how does anyone hold it to account later.”

Fifth, the honest caveat, which is strong this week and should be stated before the analysis is mistaken for a bill of indictment. CAICT has published a governance architecture that most Western standards consortia would recognise and approve of, one that explicitly refuses centralisation, distinguishes entitlement from scope, requires revocability and traceability, and anticipates reciprocity across independently governed domains. The stated limitation in §27.1 — one root domain, cross-domain recognition unspecified — is the same limitation every federation begins with, including those that later became international standards. DeepSeek has built a capital structure that solves a genuine and hard problem — extracting a large amount of patient capital while preserving founder control in a market where the alternative was not raising at all — and has done so with disclosure. The finding is not that these are failures. It is that both are the first draft of institutions whose value depends entirely on what happens in the second draft — whether the tests get published, whether the second root gets recognised, whether the post-IPO licence keeps its terms, whether a stranger ever gets a commit merged.

Sixth, what to watch, all of it checkable within a quarter. A TRADEMARK.md and a runnable conformance test suite in the OAN organisation — the two documents that would turn an accreditation scheme into a standard. The first outside contribution to OAN, by author and by review — the only measure of whether publication became distribution. DeepSeek’s prospectus, and specifically its treatment of the model licence as a policy and a risk rather than a fact — the first place a Chinese open-weight company is legally required to describe what openness commits it to. And the 2026-12-26 expiry of the first secure-and-reliable batch, which this series flagged yesterday and which falls inside the same quarter. Four documents, four institutions, and not one of them requires a policy change to observe. That is the whole argument for measuring institutions by their artefacts rather than their announcements.

Sources: 中国信通院 CAICT via WeChat — 中国信通院金键:智能体标识与身份系统是AI关键基础设施 (2026-11-20) · OpenAgenet / OAN White Paper — arXiv:2606.03161v8 (2026-11-12) · GitHub — OpenAgenet organisation, licence inventory and commit history (verified 2026-11-21) · npm registry — @openagenet/oan-sdk-ts download statistics (2026-11-21 query) · Caixin Global — DeepSeek Closes 100 Billion Yuan Funding Round (2026-11-20) · 财联社 — 9家券商参投DeepSeek,对应出资95亿,中信81亿占大头 (2026-11-19) · This series’ 2026-11-21 briefing — the bookkeeping keeps arriving: LoongArch’s twenty-two members and six commits · This series’ 2026-11-20 briefing — a toll booth in a licence · This series’ 2026-11-07 briefing — the identity code becomes a condition of market access