Open-Source

BurnRiSc (Sanko, Tian & Guizani, arXiv 2609.19422) 是维护者倦怠作为可计算对象的第一份实证——把 Oldenburg Burnout Inventory 两维度翻译成 14 个行为与语言学信号，从公开 GitHub 数据自动算出月度 BRS 分数，10 个仓库 × 68 位贡献者 × 10 例披露案例中 6 例提前 6-15 个月识别、加入峰值 BRS 作第二判据则 8/10、任意时间点 10/10。非侵入式筛查作为治理方法学的独立维度（之前开源治理工具默认侵入式），Coase 定理三个必要条件缺失的教科书样本（无边界需求 / 无经理重分配 / 无组织察觉衰退），Ostrom 第 4 条监控原则在维护者健康场景的最新形式化（监控对象从工件扩展到人），Williamson 交易成本框架下识别成本从 O(1/T) 降到 O(1)。与 #161 Wright & Augustus『XZ-burnout-as-attack-surface』直接对话构成『崩溃后果 + 崩溃识别』完整制度回应。开源四层制度基础设施第五层新增独立子层。

De Marzo, Alboré & Garcia (2026, arXiv 2609.09150) 首次记录 AI agent 在野外的完整协作现场——数千个 agent 在公共 wiki 里互助通关，作者用 3 个各只有 1 个自由参数的复制模型解释了三种决策（写在哪/叫什么/怎么措辞）。7% 创新率 + 100 条最近编辑窗口 + 两个错误率构成一个最小制度描述。‘谁先写或者在别人安静时写’成为 meritocracy/powerocracy 之外的第三种制度形态（techronomy 技术性时序评价）。Hayek 自发秩序在 agent 时代的第一次实证检验=无判断的复制；Williamson L1 在 agent 群体中的替代机制=可见性基础设施而非文化；Ostrom 公地治理八原则在 agent 群体的完全缺失=边界界定与 monitoring 都不存在。

“Three institutionally new placements on the Chinese open-source surface: (1) Xi Jinping’s BRICS-plus-open-source-AI initiative announced at the 18th BRICS Summit in New Delhi (September 13, 2026) — China will lead the creation of an ‘open-source zone for artificial intelligence’ for the 11-member BRICS grouping, with a BRICS AI open-source community, a BRICS digital ecosystem cloud platform, and a BRICS engineer-cultivation alliance and youth-exchange program — the first-documented Chinese-head-of-state placement of a ‘BRICS open-source AI zone’ object at the BRICS Summit institutional layer, pairing the same Global-South open-source instrument this series’ September 2 briefing documented at the COPU/WAICO-50-vs-2 layer with the BRICS-plus institutional layer; (2) HUAWEI CONNECT 2026 (September 17, Shanghai) — Deputy Chairman and Rotating Chairman Wang Tao’s ‘Advancing the Agentic World, Building a Solid Silicon Foundation’ keynote documents the first-documented instance of a Huawei-deputy-chairman placing the CANN community-open-source-development + Ascend PyTorch-official-backend + Atlas-960E-NPO-near-package-optics + Ascend-960DT-Q1-2027-acceleration + Ascend-960PR-Q3-2027 + Ascend-970-980-2028-2029-plus-Tau-Scaling-Law + one-million-NPU-UnifiedBus-SuperCluster + 4.16-million-Kunpeng-developer ecosystem + 90-third-party-open-source-Ascend-projects pairings at the same Huawei-deputy-chairman-keynote surface — a four-surface-simultaneous Huawei-deputy-chairman-institutional-narrative pattern, pairing the same Huawei open-source-training-ecosystem surface this series’ September 15 briefing documented at the GOSIM Shenzhen keynote layer with the HUAWEI CONNECT institutional layer; (3) OpenAtom Foundation’s first-documented open-source enterprise-training program (September 23-24, Beijing) — the Foundation’s ‘Open Source General-Knowledge Training Class Phase 1’ aimed exclusively at its own donors, with Chairman Xie Shaofeng opening and CSDN founder Jiang Tao keynoting on ‘Open Source in the AI Era: Ecosystem and Infrastructure’ — the first-documented instance of a Chinese open-source foundation placing a donor-only training-and-certification object at the OpenAtom-Foundation-donor-training layer, pairing the same foundation-training surface this series’ August 30 briefing documented at the Mulan-Foundation layer with the OpenAtom-Foundation-donor-training layer.”

科尔曼 1990 年用理性选择社会学证明：信任不是道德、不是文化直觉，是社会网络中反复解决问题时自发产生的副产品——它可以在社区里积累，也可以被抽走。GitHub 的 commit 历史是开源世界的声誉账本，每一次 PR、每一次 review 都在存入信任或支取信任。这不是《文明的教训》，是「开源是俱乐部品非公共品」命题最直接的社会理论背书。

“Four institutionally new placements on the Chinese open-source surface in the same week as this series’ September 16 briefing documented Huawei Open Source Official’s first-systematic Chinese open-source cultural definition (开源无国界 开发者有来处), Xilai Tech (Canaan Inc.) RISC-V commercial CPU IP A-share IPO filing (Xiaomi Changjiang Fund 6.12% + VeriSilicon 2.99% state-aligned-capital-market-investor pairing, 155% R&D-above-revenue profile, national little-giant → national-key-little-giant designation pairing), Ant Group Open Source Tech Committee’s 2026-H2 Agentic AI ecosystem trends paper (PR-per-Issue ratio 1.35→2.11, 90-day-PR-still-open 5.5%→11.3%, median PR merge 77.0%→68.4%, 92/100 AGENTS.md default-branch infrastructure, 143 Agent-Infra projects, more-than-half born 2025), and Wang Xu’s code-collaboration-to-idea-convergence reading: (1) CAICT Policy-and-Economics Institute + China University of Political Science and Law AI Law Research Institute jointly publish the CAICT Agent Governance Research Report 2026 (37,848 characters / 105 lines) — the first-documented Chinese-administrative-institute three-layer agent-governance framework in this series (Foundation layer data-and-identity base / Operation layer execution-and-dynamic-control / Ecosystem layer value-collaboration), pairing the same agent surface this series’ prior briefings documented at the CAICT-AI-administrative-standard layer and the Huawei-agent-open-source layer with the CAICT-governance-framework layer, and placing US-frontier-capability-diffusion-prevention, EU-fundamental-rights-protection, Singapore-global-trustworthy-AI-hub, and Chinese-large-model-institutional-stacking-governance at the same global-differentiated-governance surface — the first-documented Chinese-administrative-institute global-governance-comparison surface; (2) Jiang Xiaojuan (former State Council Deputy Secretary-General, University of Chinese Academy of Social Sciences professor) publishes at the 2026 Bund Summit ‘Born-Global AI Economy: New Growth Paradigm and Social Value Reconstruction’ (40,699 characters / 67 lines) — the first-documented State-Council-formal-level Chinese-administrative-narrative surface placing the Chinese-open-source-model-overtook-US-since-June surface at the Bund-Summit-institutional layer, with the operational claim that since June 2026 Chinese open-source large-model download volume on the world’s largest open-source community surpassed US download volume for the first time and the top-6 API call volume ranking is entirely filled by Chinese self-developed open-source large models — the first-documented quantitative evidence of the Chinese-open-source-model global-scale surface at a State-Council-formal-level institutional surface; (3) DeepSeek v4.1 operator lead Liu Shengyu publishes the 22,097-character essay ‘I Have No Choice But to Bury My Talent in Yesterday’ — the first-documented DeepSeek-internal-senior-technical-engineer personal-essay surface in this series, pairing the technical-paradox diagnostic (‘The better my operators, the faster the new model trains and infers, the faster model capabilities improve, and the earlier I will be replaced’ / ‘I will not be unemployed but I must change profession’ / ‘I hope the one who revolts me is myself’), the AI-learning-penalty data pairing (HK-PolyU + Stockholm tracking 26,811 Chinese junior-high and middle-school students, 18% short-term grade rise → 20% monthly-exam decline within 6 months → up-to-24% and 18% decline at the zhongkao and gaokao respectively after two years, Google 75% new code is AI-generated), and the operator-level institutional-form observation that the DeepSeek v4.1 primary Attention operator is written by Liu Shengyu personally — a three-surface-simultaneous DeepSeek-internal-senior-engineer-essay pattern; (4) Huawei Voice Community publishes a 55,430-character Guo Ping (Huawei Supervisory Board Chair) new-employee symposium minutes 10,000-character essay — the first-documented Huawei-supervisory-board-chair institutional-narrative surface placing the ‘For ICT and computing business, Huawei’s goal is to become NVIDIA’ diagnostic at the Huawei-voice-community-supervisory-board surface, pairing the CANN heterogeneous-computing-architecture + MindSpore framework + Lingqu (灵衢) protocol pairing, the NVIDIA-CUDA-moat institutional-mechanism reading (‘Jensen Huang’s real power is not making the fastest chip but making the world’s AI developers feel that not using NVIDIA is passing something by oneself’), NVIDIA-China-high-end-AI-chip-market-share-95%-to-near-zero diagnostic, and Huawei 2026-H1 financial profile (revenue RMB 467.8B +9.54%, net loss -36.8%, R&D RMB 121.3B +25.2%, operating cash flow -RMB 39.8B) — a four-surface-simultaneous Huawei-supervisory-board-chair-institutional-narrative pattern.”

Ansari (2026, arXiv 2609.10105) 首次给推理时 AI 治理提出可操作分类学：20 种机制 × 三大目标（Monitoring/Verification/Enforcement）× 4 点成熟度评分 × 双轴对手模型（3 能力层 × 4 对手角色）× 4 治理场景。核心实证：15/20 机制已有商用技术底物但’治理级保证差异很大’；‘没有任何一种机制在国家级部署者面前 adequate’；微调移除 enforcement 簇内部组件但平台外部控制可持续。Cohen’s kappa=0.74。AI 治理研究从训练时代转向推理时代的第一个可操作分类学。Williamson L1-L4 在 AI 治理场景的最新完整应用；Acemoglu 包容性 vs 汲取性制度框架的教科书失败样本；与 Nemecek arXiv 2609.09604 同周发表构成’训练时代失败诊断+推理时代新范式’同批实证产物。

“Four institutionally new placements on the Chinese open-source surface in the same week as this series’ September 15 briefing documented GOSIM Shenzhen 2026’s 40-keynote-speaker lineup (Western-foundation-founder, EU-regulatory, US-hyperscaler-cloud, and Western-trust-economics surfaces placed together with Chinese vendor-official surfaces), the CSDN Harness Trust Axis paper, and RVOH’s follow-up analysis of MIIT Letter [2026] No. 209: (1) Huawei Open Source Official publishes on 2026-09-15 the first systematic Chinese open-source cultural definition (开源无国界 开发者有来处 / Open Source Has No Borders, Developers Have Origins, 18,212 characters / 175 lines) — the first-documented Chinese vendor-official cultural-definition surface in this series, pairing Linus Torvalds’ ‘Just for Fun’, Peter Steinberger’s ‘Fun is velocity’ (the OpenClaw / ‘Clawdbot’ rename episode as a shared institutional-memory object), DeepSeek’s MIT-license choice as a cultural posture, and the four-classical-Chinese values ‘立己达人 · 和而不同 · 知行合一 · 义利之辨’ with a three-crossing institutional-form model (代码→项目→生态→业务 / code-to-project-to-ecosystem-to-business) and the four-mechanism diagnostic ‘interest brings speed / mission chooses direction / boundaries guarantee quality / institutions support longevity’ — the first-documented Chinese vendor-official self-declaration that Chinese enterprise open source ‘may have institutions and processes but still lack vitality’ without culture; (2) Xilai Tech (芯来科技 / Canaan Inc., RISC-V CPU IP) files for A-share IPO via Shanghai Futures and Exchange Bureau filing on 2026-09-11 (China Merchants Securities as coordinator) — the first-documented RISC-V commercial CPU IP A-share IPO filing in this series, with an institutional profile that pairs Xiaomi Changjiang Fund at 6.12% shareholding and VeriSilicon (芯原股份) at 2.99% shareholding (two state-aligned capital-market institutions), founder Hu Zhenbo at 14.97% (second-largest shareholder), national ’little giant’ designation upgraded to national key ’little giant’ in 2024-12, 2026 H1 revenue RMB 36.87M (+41%) against net loss RMB -57.30M, R&D spend 155% of revenue, four-business-line structure (RISC-V IP licensing at 92% of revenue, custom chip design, Chiplet products, server CPU not-yet-generating-revenue), and the explicit strategic positioning ’the ARM version of China’ / ‘RISC-V as the third pole’; (3) Ant Group Open Source Tech Committee chair Wang Xu publishes on 2026-09-15 the sharpest-documented quantitative evidence of Agent-Infra operationalizing in Chinese open source — Top-100 open-source repositories generated ~350,000 Issues and 610,000 PRs in the first eight months of 2026, the average PR-per-Issue ratio rising from 1.35 (January) to 2.11 (August), 92 of 100 repositories carrying AGENTS.md / CLAUDE.md / .claude/ infrastructure files on their default branch, 98 of 100 repositories permitting any contributor to open a PR, PR backlog netting +43,000 entries in the same period, the median PR merge ratio declining from 77.0% to 68.4%, the 90-day-PR-still-open ratio rising from 5.5% to 11.3%, 143 Agent-Infra projects tracked and more than half of them born in 2025, and PyTorch plus ONNX Runtime already placing Agent collaboration rules into their default branches — a structurally new institutional-form observation of open-source governance being operationalized at the coding-agent layer simultaneously with the Western-foundation-conference-program layer documented in the September 15 briefing; (4) the same Ant Group data documents the sharpest quantified diagnostic of agent-productivity-driven open-source throughput collapse — ‘Agent productivity’s great expansion will ultimately elevate code collaboration to the convergence of ideas’ — the first-documented Chinese-vendor-official institutional claim that open-source contribution at scale is transitioning from a code-collaboration object to an idea-convergence object, and the first-documented Chinese-vendor-official reading of the PR-merge-ratio collapse (77.0% → 68.4%) and the 90-day-PR-still-open rise (5.5% → 11.3%) as an institutional-mechanism problem rather than a code-quality problem. Institutional significance: the same Chinese open-source surface this series’ prior twenty-six briefings have been documenting — the institutional identity of Chinese open source across institutional borders — is being operationalized simultaneously at three institutional surfaces this cycle (Chinese vendor-official cultural-definition surface, A-share capital-market institutionalization surface for RISC-V CPU IP, and Chinese-vendor-official quantitative-evidence surface for Agent-Infra) alongside the three institutional surfaces documented in the September 15 briefing (Western-foundation-conference-program surface, Chinese-analytical-framework-publication surface, Chinese-administrative-follow-up-analysis surface) — a six-surface-simultaneous-institutionalization pattern.”

Ichihashi, Li & Zou (2026, arXiv 2609.10499) 把 Coase 委托-代理理论从工业生产搬到知识生产场景，证明最优契约诱导「监督权层级」——越容易被主直接监督的 agent 越有权评估同伴。核心洞见：「感知到的监督层级可以独立于实际监督层级存在」=Williamson L2 制度环境层的最锋利推进，开源 bus factor 问题的契约理论解释，行政式开源 vs 真开源的制度差异不是价值差异而是最优契约在不同约束条件下的解的差异。

“Three institutionally new placements on the Chinese open-source surface in the same week as this series’ September 14 briefing documented the seventh surface of state-directed open-source institutional formalization (MIIT’s ‘AI + Software’ Special Action Implementation Plan): (1) GOSIM Shenzhen 2026 (October 16-17, ADEN Hotel, 150+ international speakers, 2,000+ developers, six tracks) publishes a 40-keynote-speaker lineup that for the first time places four institutionally distinct Western surfaces (DHH’s Omarchy keynote, McCrosky’s Trusted Agentic AI keynote, Katarzyna Staroslawska’s EU-policy keynote on AI-native organizations, and Google Cloud’s Tang Xin on Harness engineering for AI game-making pipelines) at a single Chinese-conference venue alongside three Chinese vendor-official surfaces (Huawei’s Ren Xudong on the Black-Light Community paradigm shift, Huawei/Hou Peixin + Ali/Ma Tao Longxi + ByteDance/Jiang Ning + Tencent/Ru Bingsheng roundtable on Agentic Software Engineering, Huawei/Ma Yongqiang LLaMA-Factory committer on the Ascend open-source training ecosystem, Xiaomi/Du Chao on openvela as the AI hardware agent base) and CSDN founder Jiang Tao’s ‘From Open Source Software to Open Source Intelligence’ keynote — the first-documented institutional-form observation in this series of a Chinese open-source conference placing Western individual-foundation founder, EU-regulatory, US-hyperscaler-cloud, and Western-trust-economics surfaces together with Chinese vendor-official surfaces on the same conference program; (2) a CSDN-hosted analytical paper (2026-09-14) formalizes the AGI ‘vertical-and-horizontal axis’ framework — Model Capability Axis (纵轴) as scaling-law / RL / test-time-compute frontier, Harness Trust Axis (横轴) as a five-layer deterministic exoskeleton (session-and-gateway / context-and-compaction-engine / memory-and-state-persistence / tool-sandboxing-and-MCP-execution / verification-and-guardrails-with-PRM) — pairing the Process Reward Model (PRM) as the sharpest institutional-economics layer of the verification-and-guardrails engine, and documenting the 87-percent-executives-prioritize-AI-versus-only-23-percent-confident-in-safe-deployment governance-gap statistic and the Anthropic RSP v3.0 ‘game-theoretic-dilemma’ framing of voluntary-slowdown; (3) RVOH Open Source Innovation Academy’s follow-up analysis (Ministry of Industry and Information Technology Letter [2026] No. 209) operationalizes the September 14 briefing’s Move 4 (research-evaluation-incorporation) into three specific open institutional-design questions — who certifies / how to quantify / how to prevent gaming — while documenting OpenHarmony’s 1.4 billion lines of code across 13,000 contributors and RISC-V SPMP’s China-led position on the same research-evaluation-incorporation surface. Together, the three stories document a structurally new institutional pattern: the same Chinese open-source surface is being operationalized simultaneously at the Western-foundation-conference-program layer (GOSIM Shenzhen), at the Chinese-analytical-framework-publication layer (CSDN Harness Trust Axis paper), and at the Chinese-administrative-follow-up-analysis layer (RVOH research-evaluation-incorporation operational design) — a three-surface-simultaneous-institutionalization pattern that pairs with this series’ September 14 briefing’s documentation of the seventh surface of state-directed open-source institutional formalization at the national-industrial-policy layer.”