Chandana 等（2026, arXiv 2609.09519, OSCAR @ ISCA 2026）用一个 LLM agent 从 26 个开源仓库批量抽取 HLS 设计，271 候选 → 130 通过。开源代码库的角色正从「贡献者协作平台」位移为「AI agent 数据供给源」——社群完全缺席于这条流水线，派生价值归属处于制度真空。Coase 产权界定理论在 AI 训练原料场景的第一份工程实证；开源四层制度基础设施第五层向「AI 训练数据供给」方向的首次扩展。
Four institutionally dense placements on the same unsolved institutional object this series’ prior thirty-five briefings have been documenting — the institutional identity of Chinese open source across institutional borders. (1) Reuters (September 22, 2026): DeepSeek and Moonshot invited to brief the UN Security Council on AI risk this week alongside OpenAI’s Sam Altman and Anthropic senior representatives — the first-documented instance of a Chinese frontier AI lab being invited to speak at a US-led international governance forum on AI risk, with DeepSeek founder Liang Wenfeng declining attendance and Reuters documenting that Chinese frontier-lab leaders have rarely publicly articulated the possibility that increasingly capable AI systems could deceive, evade control, or cause catastrophic harm. (2) Yahoo Finance / Briefs (September 22, 2026): Microsoft considering replacing ChatGPT and Claude with Moonshot’s Kimi, and Moonshot seeking additional Nvidia Blackwell chips for its next model — a first-documented triple-Western-infrastructure-layer operationalization of a Chinese-frontier-lab-open-weight-model surface. (3) Two WeChat essays published within the same 24-hour window (September 22–23, 2026): Liu Xiangan, President of the National Institute of Electronics Standards, explicitly promoting the world’s first OSI-internationally-recognized Chinese-English bilingual Mulan Soft License and the CAICT Vice President Wang Zhiqin’s four-principles-plus-four-agent-layered-genealogy-plus-four-policy-instrument-plus-CAICT-2026-survey-40-percent-adoption-rate interpretation — the first-documented dual-state-affiliated-interpretation-layer operationalization of a single MIIT central administrative plan. (4) openKylin community’s September 22 steering committee held at the National Industrial Cybersecurity Development and Research Center (NICS), which joined as an openKylin Special-Donation Unit, with NICS co-drafting the T/SIA 056-2025 Controllable-Open-Source-OS-System Construction Specification group standard, openKylin 3.0’s kernel-jump to intelligent-native, and openKylin’s inclusion on the Chinese Foreign Ministry’s Shanghai Cooperation Organization achievement list — a first-documented five-surface-simultaneous open-source-operating-system-community-controllable-open-source-system institutionalization pattern.
Wright & Augustus 在 ACM Queue 上把开源依赖的冰山从工程问题升级为受托责任命题——冰山理论作为 Ostrom 边界界定原则的量化实现、XZ 事件把维护者倦怠重新定性为国家级攻击的攻击面、Bloomberg Sustaining Series 用持续在场而非一次性补丁把行动的定义权下放给工程师投票。
“One institutionally dense cluster on the Chinese AI-agentic-infra-and-open-source layer this series’ prior thirty-four briefings have been documenting at multiple institutionally distinct surfaces, now operationalized at three institutionally distinct surfaces simultaneously in a single 48-hour window: (1) On September 18, 2026, the Xiamen Trustworthy Agent Interconnection Center (智能体标识与身份网络暨智能体可信互联中心 厦门) launched its trial run — the second-stop launch of the HelloAgent series of release activities after a Beijing single-point validation, co-hosted by CAICT (China Academy of Information and Communications Technology), the China Internet Society (CIS), the three major telecommunications carriers (China Mobile / China Unicom / China Telecom), Fujian Broadband Network, and the Xiamen municipal government, operationalizing the MIIT-plus-three-ministry July 2026 guidance on ’exploring an interconnection network infrastructure for intelligent agents with a unified communication protocol, data interaction standard, and interface specification’ and the MIIT August 2026 15th-Five-Year-Plan for the Information and Communications Industry’s directive to ‘construct an agent network identity system, upgrade the Internet Domain Name System, establish an Internet agent identity identification system, strengthen blockchain and electronic-certification applications in trustworthy identity authentication, and build an Internet Gateway as a foundational capability for agent registration management, identity authentication, ownership determination, addressing, and resolution’ — a first-documented instance in this series of a national-standardization apparatus being operationalized at the municipal-trial-run center layer simultaneously with a Beijing single-point-validation layer and a national-integrated-network-deployment layer, adopting an ‘agent + DID + blockchain’ fusion technology route and positioning the Internet Gateway as the ’next-generation Internet infrastructure’; (2) In the same 48-hour window, Moonshot AI’s Kimi K3 open-weight model became the first-documented case of a Chinese frontier-AI-lab open-weight model license placing a US$20 million aggregate revenue threshold and a 100 million monthly-active-users attribution requirement on Model-as-a-Service commercial deployment, with Bloomberg and Reuters reporting that Moonshot AI is targeting US$2 billion in annualized revenue by year-end 2026 (up from US$1 billion run rate in August, up from US$300 million in June, up from US$200 million in April), pursuing confidential Hong Kong IPO filings with Goldman Sachs / China International Capital Corporation / Deutsche Bank as underwriting advisors at a US$50 billion valuation, and negotiating with Microsoft / Amazon / Google for up to a 30% revenue-share hosting arrangement — the same Moonshot surface this series’ September 12 briefing documented at the SCMP dual-listing-exploration layer and this series’ September 1 briefing documented at the National AI Industry Investment Fund lead-investor layer is being operationalized at the open-weight-license-commercial-monetization layer simultaneously — a structurally new institutional-form observation on the Chinese-frontier-AI open-weight commercial-monetization layer; (3) In the same 48-hour window, the Tiangong Kaiwu Open Source Foundation’s AI Open-Source Ecosystem Operations Class — the class this series’ September 20 briefing documented at the Chinese municipal-district-party-committee-talent-office-plus-university-plus-open-source-foundation-plus-education-platform-plus-weekly-event co-hosting layer during the GOAI Open Source Week Ecosystem Co-Builder Day — published its first batch of 11 concrete ecosystem-operations tasks on September 20, 2026, with the closing assignment submission deadline at 12:00 noon September 22 and approximately 20 participants advancing to the offline Demo Day on September 23, comprising 11 tasks ranging from (01) converting university events into sustained communities to (11) helping users select and use H3 community solutions (MiniMax H3), with task (02) — building the KHub rare-disease open source community — being the only task on the list that is not positioned at the university-community, industry-publicity, industry-training, or AI-café operationalization layer but at the public-health-and-grassroots-open-source-collaboration layer, and with Datawhale’s co-hosting and Tiangong Kaiwu Foundation Deputy Secretary-General Zhuang Biao-wei’s facilitation placing the class in the same institutional lineage as the Foundation’s own diagnostic this series’ September 20 briefing documented verbatim (’technology is only the starting point; how technical value is understood; how a project continues to attract participants; how a single event can be deposited as long-term collaboration’) — a first-documented instance in this series of a Chinese open-source foundation moving from a foundation-plus-municipal-party-committee-talent-office co-hosting Ecosystem Co-Builder Day event layer to a foundation-curated-11-concrete-task-list layer, with the class’s own methodological formula (‘find real needs → find the right people → design the participation path → enable the first action → let contributions keep happening’) being a first-documented instructional-ecosystem-operator-talent object at the Chinese open-source ecosystem-operations surface. The three placements together document a structurally new three-surface-simultaneous institutional pattern on the Chinese AI-agentic-infra-and-open-source layer: the MIIT-15th-Five-Year-Plan-plus-three-ministry-guidance-plus-CAICT-plus-CIS-plus-three-major-carriers-plus-municipal-government national-standardization apparatus at the Xiamen municipal-trial-run-center layer; the Moonshot-AI-Kimi-K3 open-weight-license commercial-monetization-and-margin-gap apparatus at the Chinese-frontier-AI open-weight-commercial-monetization layer; and the Tiangong-Kaiwu-Foundation 11-real-task-list instructional-curriculum apparatus at the Chinese-open-source-ecosystem-operations-talent-cultivation layer. This briefing documents — but does not resolve — the question of whether the same Chinese open-source surface is capable of holding three institutionally distinct institutional forms simultaneously without one converting the other, and whether the ‘agent + DID + blockchain’ fusion technology route, the Moonshot open-weight-license commercial-monetization thresholds, and the 11-real-ecosystem-operations-task-list instructional curriculum represent three independently evolving institutional forms on the Chinese AI-agentic-infra-and-open-source layer, or three convergent expressions of a single state-directed-institutional-form object. The editorial perspective this briefing adopts is a Williamson L1→L4 institutional-economics reading (L1 social embedding → L2 institutional environment → L3 governance mechanisms → L4 resource allocation) cross-referenced against the Acemoglu-Robinson inclusionary-extractive institutional dichotomy, the Elinor Ostrom eight-design-principles for commons-governance, and the Great Divergence 2.0 framework distinguishing FLOSS from State-Chartered Codebase / Intranet Shared Source / Cyber-Estate.”
Brömme 提出 product- and vendor-neutral 的 Agent 过程 black-box 证据架构——六步管线把 agent 事件产权拆成『内容控制权』+『完整性证明权』。Agent 信任基础设施第五层的第一个可操作架构提案。
埃里克森 1991 年用加州沙斯塔县的牧场主田野证据证明：秩序的第一性来源不是法律，是社区规范。邻居不需要法院就能解决越界牲畜的纠纷——因为他们相邻、可观察、可报复。开源社区的维护者、review、贡献指南，本质上是同一类规范基础设施。这是「开源是俱乐部品非公共品」命题的规范学支点，也是大分流 2.0「行政动员挤出自发规范」命题最锋利的制度证据。
“Three institutionally new placements on the same unsolved institutional object that this series’ prior thirty-two briefings have been documenting: the institutional identity of Chinese open source across institutional borders. (1) MIIT Vice Minister Xin Guobin, speaking at the 2026 World Manufacturing Conference in Hefei on September 20, 2026, directs MIIT to vigorously develop open-source base models and vertical-domain models to consolidate the software and hardware foundation of new-generation intelligent manufacturing - the first-documented instance of a MIIT deputy-minister-level official directly naming open-source base models as a state-directed industrial-policy surface in this series, and the first-documented administrative-level operationalization of a term that no open-source community has yet defined; (2) OpenAtom Foundation Chairman Xie Shaofeng delivers the keynote Open-Source Innovation Building the Software Powerhouse City at the 28th China International Software Expo in Zhengzhou on September 20, 2026, framing open source as the state-directed instrument for late-developing cities to swap-lane-and-overtake the 14 existing national software powerhouse cities, co-hosting three work items (build famous enterprises / build famous parks / cultivate talent) with a government-plus-foundation-plus-local-service-system tri-party operational structure; (3) The Supreme People’s Court Opinions on AI Disputes (September 7, 2026) explicitly names open-source software developers in Article XIII for liability exemption - the first-documented Chinese-judicial apparatus placing open-source developers as a distinct legal subject for tort liability - while conspicuously sidestepping the copyrightability of AI-generated works themselves.”
After the Party (Xiong & Zhang, arXiv 2609.17274, APSEC 2026) 是 agent skill registry 作为独立治理对象的第一份实证——6 万 skill / 91 天近乎翻倍 / Top 10% 拿走 46.93% 下载 / 77.86% 零 star 零评论 / 85.06% 特权证据 / 三个扫描器在 61,990 skill 上 23,702 处意见不合 / 加权敏感性仅 21.67%–61.06%。『自然语言指令 = 新代码』命题的第一份量化实证，治理真空作为既非行政动员也非自发秩序的第三种状态，合规审计制度供给 < 制度需求多源实证群的第七份样本（首个行为合规样本），『派对结束后』作为治理研究时序视角的独立维度。开源四层制度基础设施第五层第四次扩展：贡献政策 → 信任基础设施 → 推理时治理 → agent skill 治理。
“Two institutionally new placements on the Chinese open-source surface this series’ prior thirty-one briefings have been documenting at multiple layers, now documented together for the first time at the Chinese municipal-district-government-talent-office-plus-university-plus-open-source-foundation-plus-education-platform-plus-weekly-event co-hosting layer and the CAICT-administrative-institute three-tier-agent-assessment-instrument layer: (1) The Tiangong Kaiwu Open Source Foundation (天工开物开源基金会), co-hosted with the Xihu District Committee Talent Office (西湖区委组织部人才办), Hangzhou Dianzi University (杭州电子科技大学), and Datawhale, launches China’s first AI open-source ecosystem operation class (全国首个人工智能开源生态运营班) — the first-documented Chinese municipal-district-party-committee-talent-office-plus-university-plus-open-source-foundation-plus-education-platform plus weekly-event co-hosting surface in this series, launching the inaugural Ecosystem Co-Builder Day (首届生态共建者日) on September 20, 2026 during the GOAI Open Source Week, and pairing the operational claim ’let talent grow in projects; let projects be active because of talent’ with the Foundation’s own diagnostic that ’technical value must be understood by projects that can attract participants and by events that can be deposited as long-term collaboration’ — the first-documented Chinese open-source foundation pairing a party-committee-talent-office surface with a foundation-ecosystem-operations surface simultaneously; (2) CAICT (中国信息通信研究院) launches the first batch of Agent Management Ability Maturity Assessment (AOMA) — a five-level maturity framework (Entry / Advanced / Excellent / Excellence / Leading) with a two-axis management-domain architecture (comprehensive management domain + full-process management domain), jointly designed with 50+ enterprises and nearly 100 experts, positioning itself as a third-party assessment-and-testing system for AI-native organizational capability — the first-documented CAICT-administrative-institute three-tier-agent-assessment-instrument surface in this series, operationalizing the agent surface this series’ September 17 briefing documented at the CAICT-administrative-institute three-layer-governance-framework layer and this series’ September 14 briefing documented at the MIIT ‘AI + Software’ Special Action Implementation Plan’s Agent Software Application Store + Skills Resource Library layer at the CAICT-administrative-institute-assessment-instrument layer simultaneously — a two-surface-simultaneous CAICT-administrative-institute-agent-assessment-instrument pattern.”