Home › Research › Research register: every line of research, including negative results

Research register: every line of research, including negative results

Source file: docs/RESEARCH-REGISTER.md · updated 2026-11-08

Part of the methodology (docs/METHODOLOGY.md §3.2). Started 3 Nov 2026, backfilled for work since 29 Oct 2026.

Each line of research is a branch that stays open over time. An entry records what was asked, how, what worked, what did not work, and how to re-run it. Later updates are appended as dated rounds under the same ID, never overwritten, so the method itself can be repeated, audited and continued by anyone.

Outcome codes: ✔ conclusive · ◐ partial · ○ inconclusive · ✖ negative (searched, nothing found — itself a result) · ⊘ blocked (access failed). Depth: 0–4 as in docs/RESEARCH-DEPTH.md.

Index

R04 — PA reuse of open source

• Question: which public administrations actually run open-source software (not plans), EU vs CN, open source only (HarmonyOS excluded, Fabio 2 Nov)?
• Rounds: R1–R3 (31 Oct–1 Nov, web + agents), R4–R6 (2 Nov, Qwen/DeepSeek leads + server verification).
• Method: official procurement notices, e-gov cloud catalogues, price lists, final accounts; leads from Chinese LLMs verified with curl + pdftotext + OCR on the server.
• Worked: e-gov cloud service catalogues and procurement specs (Fujian, Yancheng, Jinan, Shijiazhuang) — the pattern "openEuler as OS option / requirement" is the productive search key.
• Did not work / negative:
• ✖ No Chinese government code-sharing platform (confirmed by DeepSeek and Qwen independently).
• ○ Hulunbuir e-gov cloud contract: Qwen quoted "华为欧拉"; OCR of 34 pages found only Kylin/UOS → claim rejected.
• ○ Xuancheng lead: not checkable.
• ◐ Changde: only security services supporting openEuler guests → context, not counted.
• Not yet checked: Shanghai Fengxian, Pingli county, OpenHarmony water-monitoring terminals (Qwen round 6).
• Re-run recipe: for each of the 31 provinces search "政务云" "服务目录" openEuler OR 欧拉 OR 龙蜥 OR OpenCloudOS site:gov.cn, plus 采购需求 "操作系统" openEuler; verify on the PDF; record negatives here.

R05 — EU–China bridges

• Question: cooperation with official sources on both sides naming each other.
• Worked: Eclipse–OpenAtom (Oniro); EU–China High-Level Digital Dialogue (channel; open source not on the agenda); Catena-X ↔ CAAM/VDA/CAICT (ChatGPT for EU side, DeepSeek for CN side; CAAM pages + Jiangsu Data Bureau).
• Did not work: ✖ CORDIS 101137971 (ChatGPT marked it as a bridge; it is a decarbonisation project). ◐ EPL-2.0 Chinese translation: OpenAtom page found, no Eclipse-side source. Borderline: OpenChain ISO 18974 (international standard, not a bilateral bridge).
• Lesson: LLM "verified" flags are unreliable; require one official page per side.
• Re-run recipe: for each EU flagship (Eclipse, Gaia-X, Catena-X, OpenInfra Europe, OW2) search its newsroom for China/Chinese partners and the Chinese partner's site for the EU name.

R08 — Chinese originals without VPN

• Routes tried, in order, with results (3 Nov): server (Paris datacenter IP): many .gov.cn time out or 403. Local residential IP (Italy): some work (Jiangsu Data Bureau 200 where server got 403). Wayback Machine: worked for the Shenzhen gazette (id_ + --compressed required, gzip otherwise). archive.today: not yet tried.
• ⊘ Still blocked: cq.gov.cn ×2 (Chongqing), lg.gov.cn (Longgang), szlhq.gov.cn (Longhua) — no Wayback copy.
• Lesson: "needs VPN" is a last resort; always try residential IP → server → Wayback → reprints → LLM quotation (A-v).

R09 — Chinese OSS landscape

• Method: GitHub/Gitee/OpenDigger APIs (measured, A), China Open Source Annual Report 2025 via press relay (B).
• Gaps / inconclusive: Gitee GVP full list, AtomGit G-Star list, CAICT and OSCHINA reports not fetched; platform user counts are self-reported (C) and not comparable with GitHub active developers.

R10 — COSCon themes

• Worked: themes 2018–2025 verified (2018 from kaiyuanshe.github.io Activity archive).
• Inconclusive: 2016, 2017 themes only from Qwen (D); full track lists missing most years; 2020 city, 2023 dates to confirm.
• Re-run recipe: kaiyuanshe.github.io/categories/Activity pages + Wayback of kaiyuanshe.cn; one entry per new edition each December.

R11 — Chinese foundations, OpenAtom

• Worked: OpenAtom statutory annual work reports 2020–2023 from its disclosure page → scanned PDFs → OCR (pdftoppm -r 200 + tesseract chi_sim+eng on the server) → income, spending, donors, government subsidy (¥20.7M in 2023).
• Did not work: ⊘ Ministry of Civil Affairs registries (chinanpo, 慈善中国) are JS-only, unreachable from our hosts. ✖ 2024–2025 reports not published. ○ Xinhua "¥600M+ since 2020" cannot be reconciled with filed ¥270M (2020–23) → kept as C. Licence check: 35 repos OSI-approved, ~15 unchecked (incl. AI models).
• Caveat: OCR figures must be re-read on page images before publication.

R12 — Public funding (amounts)

• Rules adopted during the branch (Fabio, 3 Nov): open-source-specific instruments only; generic innovation/R&D tax credits excluded and logged in data/funding/excluded.json; adoption subsidies and grants in scope; EUR at ECB annual average.
• Worked: EU programme envelopes (CORDIS), German STF budgets (press, B), Chinese local caps (Shenzhen A via Wayback, Hangzhou, Guangdong).
• Negative / inconclusive: ✖ no central Chinese open-source budget line (MIIT 揭榜挂帅 2026 batch: no OSS topic; 超长期特别国债 not linked to 开源鸿蒙); ✖ 23 EU member states "none found" after light search only; ○ no Chinese disbursements (award notices 拟资助名单 not collected); OSOR, EU-FOSSA, IPCEI-CIS public share not found.
• Re-run recipe: yearly, after budgets are adopted: Bundeshaushalt Einzelplan 09 (STF), CORDIS API for NGI/OSS calls, Chinese award notices per city.

R14 — Private corporate funding

• Worked: FSFE, KDE e.V. (filed corporate lines), Eclipse annual reports (total revenue only), OpenAtom donations.
• Inconclusive: Eclipse revenue not split between member fees and public grants; NLnet private share; OW2, LF Europe, company FOSS funds not quantified; EU Transparency Register not yet consulted.

R16 — Multi-LLM iterative research (cross-cutting)

• Setup: ChatGPT Plus, DeepSeek free, Qwen free, via the user's logged-in Chrome; no API. Runs in reviews/ + data/llm-runs.tsv.
• Observed behaviour: Qwen best on .gov.cn documents but misquoted once (Hulunbuir); DeepSeek loses web search on follow-up turns (open a new chat per question); ChatGPT "verified" flags unreliable; ChatGPT home page swallowed input twice (send from an existing chat or via JS on the composer).
• Operational lessons: strip query strings before reading output; keep waits ≤ 40 s; ask for compact lists (output read in ~1k-character slices).

(R01–R03, R06, R07, R13, R15: see their output files; entries to be expanded at the next round of each branch.)

R15 — Funding of major projects

• Round 1 (3 Nov): 34 projects (18 CN-origin, 16 EU). Method: project/foundation reports, Open Collective, STF and CORDIS records, OpenDigger (Doris only).
• Worked: STF per-project grants, Open Collective totals, commercial models.
• Inconclusive: employer shares of top contributors mostly inferred (B/C); SCS federal grant (~€15M), GnuPG STF amount, Matrix.org accounts to confirm.
• Incident: an agent ran a stale scratch script that overwrote data/initiatives.json and data/new_source_urls.txt; both restored from git, verified identical to HEAD. Lesson: agents must not run scripts found in the scratchpad.
• Re-run recipe: OpenDigger developer/affiliation endpoints per repo; Open Collective API; STF "Technologies" page; foundation annual reports.

Chapter closures

• Funding (R12–R15) → docs/FUNDING-CHAPTER.md v1.0, 3 Nov 2026.

Round 3 Nov 2026 (evening) — free-tier harvest, Qwen + DeepSeek (Fabio: "keep them working until the free tier ends")

• R12: first Chinese award-stage records (D, verification blocked): Shenzhen 2024 batch 2 重大开源项目 拟资助 4 projects ¥8.32M (gxj.sz.gov.cn ⊘ TLS); Hangzhou Xihu AI-open-source-community compute voucher to 通义云启 ¥26.85M (WeChat relay only); Guangdong 2026 OpenHarmony adaptation centre ¥10M to 科学城（广州）数字科技集团 (relay mziit.com fetched, B pending table check). ✖ Wuhan East Lake; ✖ Shanghai AI open-source award list not public; ✖ Shenzhen 2025–26 results not yet published. New instrument type to model: OSS-community compute vouchers (R13).
• R11: ✖ OpenAtom 2024–25 financials still unpublished; 2026 new platinum/gold donors listed (D).
• R09: Gitee GVP 466+, AtomGit G-Star 1,111 joined / 577 graduated (C, self-reported).
• R10: COSCon 2016 slogan conflict (DeepSeek "天下开源是一家、社区携手迸火花" vs Qwen "源力觉醒") → both D; 2017 Shanghai SJTU 18–19 Nov; annual report URLs 2022–25 collected.
• R04: ✖ OpenHarmony water terminals (Qwen withdrew its lead); Fengxian tender verified on PDF but openEuler only in antivirus compatibility → context; Pingli: compatibility clause → context. Lesson added: compatibility ≠ adoption.
• R16: Qwen search stalls → "Salta" after ~60 s forces an answer. Neither free tier hit a limit in 10 prompts; harvest continues in the daily task.
• R12 (cont.): ✔ Wuhan 2024 excellent open-source projects and communities — first Chinese award-stage list verified on the original (A, archived): 10 projects + 1 community, amounts not published. Beijing city caps ¥2M/¥1M (guide). Jiangsu, Zhejiang, Chengdu ✖ no lists.
• R08 (VPN queue): docs/CN-VPN-QUEUE.md now lists 31 queued items + all 288 distinct Chinese URLs cited in the repo; tools/vpn_fetch.sh --vpn-on makes a local copy (with attachments 公示表/PDF/XLS, SHA-256) into sources/archive-vpn/. Incident: an unescaped backtick in vpn_queue.sh executed vpn_fetch.sh without VPN; stopped, output deleted, script now refuses to run without --vpn-on.

Round 4 Nov 2026 — daily task (no VPN)

• R08 ◐ (reachability). Question: which cited Chinese URLs really need a VPN? Method: the output of the 3 Nov accidental no-VPN run of tools/vpn_fetch.sh (from a European residential IP) had been appended to CN-VPN-QUEUE.md. It is now moved to data/cn_reachability_noVPN_2026-11-03.tsv and the doc is restored. Result (fact): 284 URLs, 164 → 200, 101 → 000 (timeout), 7 → 521 (CSDN, Cloudflare), 6 → 404, 2 → 412 (caict.ac.cn), 2 → 418 (gitcode.net anti-bot). Re-test today: gxj.sz.gov.cn answers the laptop but not the Paris server (000), and gdii.gd.gov.cn does the opposite. Lesson: always try both routes before queuing a URL as "VPN only". Re-run: curl -s -o /dev/null -w '%{http_code}' -m 20 -A 'Mozilla/5.0' URL from the laptop and from ssh ubuntu@51.158.36.151, plus the Wayback availability API.
• R08 ✔ (4 queue items closed). Fetched directly from the laptop and archived with SHA-256: the Shenzhen 2024 重大开源项目 award notice + 公示表 .xls, the Shenzhen 2024 application guide (zip with 2 guides, the audit rules and an operating guide), the Longhua 2026 鸿蒙 e-government matchmaking report (区工信局, 27 May 2026) and Jiemian's 矿鸿 launch article. Longhua is HarmonyOS/开源鸿蒙 ecosystem promotion to e-government suppliers. It is a matchmaking event, not a deployment, so it counts as context. Still queued: 17 lines in data/cn_vpn_needed.txt.
• R12 ✔ (first CN award with amounts on the original). Shenzhen MIIT bureau, 25 Nov 2024, 2024 batch 2 重大开源项目 拟资助: 4 firms, ¥8.32M (A). Beneficiaries: 鸿元智通 Zhihongmeta OS distro ¥1.51M; 证通电子 OpenHarmony finance/e-gov IoT terminals ¥2.79M; 黑蜂智造 smart ear-cleaning device ¥2.45M; 前海联大 cloud speaker ¥1.57M. Guide (A): "major open-source project" = donated to OpenAtom by a Shenzhen company; terminal OS >300M devices, or server OS whose donor is in the Linux kernel changeset top list (inference: OpenHarmony and openEuler, both from Huawei). Rates: distro 30% of audited sales; chip modules ¥10/40/150 per device at 50/40/30% over 3 years. Finding (inference): Shenzhen's open-source money rewards downstream adoption and shipments, not upstream development, and no code release is required. data/funding/cn.json: 2 records C→A, 1 new record (proposed-award, not summable: batch 1 not found). Side note: the .xls also carries an unrelated sheet (首版次软件 audit, 27 firms), which is out of scope. Next: Shenzhen 2024 batch 1 and 2025–26 公示 on gxj.sz.gov.cn (reachable from the laptop); Guangdong ¥10M original via gdii.gd.gov.cn from the server.
• R02 ◐ (academic databases, first systematic pass). OpenAlex API, 9 queries, 2019+, top 12 by citations each (tools/openalex_query.js; raw output in data/lit/openalex_2026-11-04.md). Hits: 77, 73, 137, 152, 21, 1260, 108, 70 and 1. Screening: ~100 records seen, 7 retained, 0 that compare EU and Chinese open-source policy. Negative (✖ for the comparison question): OpenAlex abstract search is noisy for "open source" (it matches OSINT, open data and medical AI), and Chinese-term search (开源 政策) does not tokenise and returns ecology papers. Not done: Google Scholar, SSRN, Semantic Scholar, CNKI. Re-run: node tools/openalex_query.js after editing the qs array, and prefer title.search.

Round 5 Nov 2026 — daily task (no VPN)

• R12 ⊘ ✖ (Shenzhen 2024 batch 1 and 2025–26 重大开源项目 公示). Question: do the other Shenzhen award batches exist, and with which amounts? Routes: (1) direct curl from the laptop to gxj.sz.gov.cn (which worked on 4 Nov) → 000 timeout, including the already-archived URL, over http and https; (2) Paris server → 000; (3) Wayback CDX gxj.sz.gov.cn/gkmlpt/content/ from 2024 → 129 captures (14 from 2024, 106 from 2025, 9 from 2026), titles fetched one by one (124 read, 5 failed): none is an open-source award or guide (only one software-sector notice, a 2023 Q4 output bonus); (4) Qwen (see R16) → fabricated URLs, rejected. Outcome: blocked today, negative on Wayback. Re-run: retry the laptop route on another day (the site is intermittent); CDX query also on gxj.sz.gov.cn/xxgk/xxgkml/qt/tzgg/; the gkmlpt API /gkmlpt/api/all/0?page=N once the site answers.
• R04 / R06 ◐ (OpenCloudOS / OpenAnolis in government). Question: is any named government body or e-government cloud running OpenCloudOS or Anolis OS (deployment or purchase, not compatibility)? Method: DeepSeek (DeepThink+Search, new chat), then Claude verification on the original. Result: ✔ Shenfu Reform and Innovation Demonstration Zone (Liaoning) XC e-government cloud, phase I 50 Loongson servers all on Anolis OS 8.9 (fact as reported by the OpenAnolis community case study, 22 Nov 2024; B, archived with a verbatim extract because the article body is loaded by JavaScript). Context: 政采云 (CentOS → Anolis 8.2) is a company, not a government body. ✖ No named OpenCloudOS government user (only vendor claims). ✖ No gov.cn/ccgp tender or award naming either OS as the purchased item; hits are compatibility clauses. Catalogue: 1 new record, gate_a_counts=false until a government source confirms it. Re-run: ccgp.gov.cn / ccgp-liaoning search "沈抚 龙芯 云" for the procurement record; OpenAnolis 案例集 V1/V2 PDFs for further government cases.
• R02 ◐ (Semantic Scholar). 8 queries; only 2 returned (HTTP 429 on 6, from the laptop and the server). 30 titles screened, 4 included, 0 comparative EU–CN. Details and re-run: LITERATURE-REVIEW §5 round 2.
• R16 (LLM behaviour). DeepSeek free: correct use of the compatibility exclusion; it said "very few" instead of padding the list. Qwen free: after "Salta" it answered without retrieval and produced self-declared placeholder URLs ("示例链接结构") despite "不要编造网址" in the prompt; the post-ID format (7 digits vs the real 8) gives them away. New rule: after a Qwen "Salta", every URL is presumed unverified. Neither free tier showed a limit message (1 prompt each, kept small for budget).

Round 6 Nov 2026 — daily task (no VPN)

• R12 ✔ (Guangdong ¥10M original verified A + 2 new OSS awards). Question: verify the 3 Nov mziit.com relay lead (OpenHarmony adaptation centre ¥10M) on the original. Routes: gdii.gd.gov.cn answered the laptop today (it had been server-only); its unified search API is 403, Bing empty → crawled the 通知公告 list (zwgk/tzgg1011) pages 1–40 by title. Found and archived (A, with 4 attachments incl. the .wps amounts table parsed by UTF-16 extraction): 《关于下达2026年省级制造业当家重点任务保障专项资金（新一代信息技术和产业发展）项目计划的通知》, 2026-02-14. Amounts table (万元): 开源鸿蒙适配中心建设项目 → 科学城（广州）数字科技集团 ¥10M (补改投); 广州智算开源融合赋能中心 → 广州数据集团 ¥8M; AI大模型/智能体测试验证开源生态中心 → 广州掌动智能 ¥3.39M; one non-OSS line ¥2.8M; table total ¥24.19M. 附件3 (target sheets, A): adaptation centre = public-interest company-operated platform, full-chain adaptation, verticals incl. 政务; KPIs include open-source tool/dataset releases. data/funding/cn.json +3 records (A, summable); tables rebuilt. Negative: the 2025 edition of the same fund (电子信息 direction, post_4660522 + .xls, archived) has no OSS line (UTF-16 both offsets + GBK scan) → the OSS lines are a 2026 novelty in the 信息化与软件服务业 direction. Re-run: crawl tzgg1011 titles each February for the next 下达 notice.
• R04/R06 ◐ (Shenfu B→A attempt → operator model documented). Question: find the official procurement record naming Anolis/Loongson. Routes: ccgp.gov.cn rate-limits laptop AND server ("频繁访问"); ccgp-liaoning 000 → DeepSeek round 1: no official notice; the 政务云一期 purchaser is China Unicom Liaoning (sub-packages PAAS/cloud-desktop/cabling on third-party tender sites) → DeepSeek round 2 on the 2025-12-30 《沈抚示范区信创云平台及配套设施建设项目中标公示》: amount ¥6,495,000/yr cloud-lease to 联通抚顺分公司 (A-v via chinabidding, EdgeOne-blocked for us) + ¥120k 监理 (北方实验室); Claude fetched zhaobiao.cn with the real browser (curl/jina blocked 521/567/403) and archived a verbatim extract (B): spec requires 全面国产化替代 of OS/DB/compute/storage, 3-year lease then ownership transfer, no CPU or OS brand named. Shenfu stays B, gate_a_counts=false. Finding (inference): CN XC procurement specs are brand-neutral; the concrete OS surfaces only in vendor/community case studies — so "no ccgp record naming openEuler/Anolis" is structural, not an accident of this case. Re-run: ccgp search when un-throttled; 辽宁公共资源交易 portal for the original.
• R02 ◐ (Semantic Scholar round 3). 1/6 pending queries returned (S4); 15 screened, 4 included; cumulative 15 included, 0 EU–CN comparative. Details in LITERATURE-REVIEW §5 round 3.
• R12 ✖ (Shenzhen batches, Qwen). gxj.sz.gov.cn still 000 from laptop and server. Qwen search stalled → Salta → no-retrieval answer 未找到 batch 1/2025/26 (D), converging with our 5 Nov Wayback negative. Nothing actionable; retry the site another day.

Round 7 Nov 2026 — daily task (no VPN)

• R12 ⊘ (Shenzhen batches, 3rd attempt). gxj.sz.gov.cn (site root and /gkmlpt/api/all/0?page=1) → 000 from laptop and Paris server, third consecutive day. Re-run unchanged; consider the Wayback CDX on gxj.sz.gov.cn/xxgk/xxgkml/qt/tzgg/* next time.
• R06 ✖ (SecretFlow 隐语 in government — CN data-space layer). Question: is any named government body / public-data platform verifiably running open-source SecretFlow? Method: DeepSeek (DeepThink+Search, new chat) → Claude verification on the originals. Result: DeepSeek found none meeting the criteria (site:gov.cn and ccgp empty) and self-flagged 4 near-misses. Verified: (a) Hangzhou 密态计算中心 (2024-09-26, Hangzhou Daily via ifeng, archived, B): promoted by 杭州市数据资源局 + 发改委, built with Ant's 蚂蚁密算 — text says "密态计算技术体系", not 隐语 nor open source → context; (b) Wuhan gov site whhlwdj.gov.cn: 隐语开源社区武汉站 launched in Optics Valley (A for the event, archived) → community, not deployment, context; (c) SecretFlow case page (Shandong 医保) JS-rendered, unreadable by curl and the real browser (⊘); (d) Jiangxi, Ma'anshan: SecretFlow not named (✖). Finding (inference, converges with R04/R06 5–6 Nov): Chinese government texts describe the technology class (隐私计算/密态计算), never the open-source project. Re-run: inspect the SecretFlow case page network calls; ybj.shandong.gov.cn "隐语"; 浙江政府采购网 "密算".
• R04/R06 ✖ (OpenAnolis 案例集 government cases, Qwen). Search stalled ~95 s → Salta → placeholder URL self-labelled "示例失效" (Gansu e-gov cloud, D) + unnamed Beijing district → rejected. Second Qwen placeholder incident after Salta (first 5 Nov). Rule refined (R16): for URL-seeking questions, abandon a stalled Qwen query instead of clicking Salta. Re-run: DeepSeek with the same prompt; openanolis.cn cases via real browser.
• R02 ◐ (literature: arXiv added as 3rd database). New keyless tool tools/arxiv_query.js; 6 queries, 117 hits, 83 screened, 10 included; cumulative 25 included, 0 EU–CN comparative policy works. Semantic Scholar S1 S2 S3 S6 S8 re-run in parallel (all 5 refused, HTTP 429 — keyless route not workable; needs a free API key or OpenAlex substitutes; see LITERATURE-REVIEW §5). Details in LITERATURE-REVIEW §5 round 4.
• R16 (free-tier status). DeepSeek 1 prompt, Qwen 1 prompt today (plus this morning's ZH copy-edit runs): no limit message on either.