AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-10

Daily gap hunt — what the wiki is missing and how each gap was triaged.

Contents

Scanned

Recent window: 6 New Developments Log/ files covering 2026-09-07 to 2026-09-09 (unprocessed_devlog=0 — the first run in nine to find the digest backlog clear), and 78 wiki pages with last_updated in the last 48 hours. Rotation slice: 0 — companies/ (all 168 pages).

bin/lint-scan.py run fresh: banned_headers=0 date_headers=26 self_refs=6 predictions=0 header_hype=0 broken_distinct=213 unprocessed_devlog=0. bin/primary-doc-scan.py --days 3: 2 cited primary-document URLs across 6 digests, 1 missing from Raw Sources/, 0 possible-match-under-another-address, 1 second-lane keyword hit.

Gaps actioned (8 of 19 found)

New pages created (live)

None, and deliberately. Every missing-page candidate this run either resolved to an existing page under another name (below) or failed the quality gate for want of first-party material. Two candidates that would have been created on merit — entities/bernie-sanders and entities/tech-transparency-project — are in the deferred backlog with the reason stated.

Alias resolution mattered more this run than in any recent one. Three candidates that a broken-link or name scan would have reported as missing pages are not:

Pages expanded (live)

  • UK Artificial Superintelligence Bill (ControlAI draft, May 28, 2026; introduced September 8, 2026) — score 8 (live thread +3, primary document +3, frontier-governance core +2), and the substantive finding of the run. The page was last_updated: 2026-06-06 and recorded the bill's status as "Awaiting introduction" through the private-members'-bill ballot. The bill was introduced in the House of Commons on September 8, 2026 by Alex Sobel MP (Labour, Leeds Central and Headingley) — and by a different route than the page anticipated, under the Ten Minute Rule rather than through the ballot the page describes ControlAI's contact tool as targeting. The page therefore asserted a stale status on an instrument that had already moved, on a live thread. Rewritten: introduction and first reading folded into the existing status-and-timeline section as prose with the ballot history preserved as prior context; parliamentary vehicle, sponsor, long title and passage state added to the header table; the definitional content and the Ten Minute Rule's low prospects added; a ## Open questions section added recording that the operative text as introduced was not published on the parliamentary record as of today, so the provisions section still rests on the ControlAI draft. sources_count 2 → 6, last_updated bumped, confidence held at medium — the introduced text is not held, which is exactly the limit medium is for. Four new (Source: URL) cites plus the parliamentary record. Nothing removed.

Two divergences recorded rather than resolved. (1) The bill has two names. ControlAI calls it the Artificial Superintelligence Security Bill; the UK Parliament's own record at bills.parliament.uk/bills/4288 titles it the Artificial Superintelligence Bill, without "Security". The vault's slug and prior title followed ControlAI. Both names are now on the page with a note on which to use for which citation. (2) The two accounts claim different firsts. TIME reports Sobel introducing "what he said was the first bill seeking to ban superintelligence in any G7 parliament"; ControlAI describes it as "the first bill to prohibit superintelligence ever introduced in a legislature anywhere in the world." The claims differ in scope and neither is asserted on the page.

Five companies/ pages at the top of the rotation slice by in-degree, all lacking a typed ## Relationships section. Each edit is purely additive graph wiring plus one last_updated value; no prose was changed, no fact or citation removed, and no confidence or sources_count raised — wiring adds no evidence.

  • Anthropic — score 4 (in-degree 819, frontier-lab core +2). The highest-in-degree page found in any slice audited to date, ahead of California SB 53 — Transparency in Frontier AI Act (254) and NIST AI Risk Management Framework (AI RMF 1.0) (163), and it had no typed relationships at all. Thirteen typed links. Three carry a claim rather than a bare pointer: litigates: Anthropic v. United States (Pentagon ban challenge) as plaintiff; contradicts: ITI — Information Technology Industry Council, stating the substance of the September 8 membership split rather than narrating it, and reciprocating the edge that page added on 2026-09-09; and supports: Responsible Scaling Policy (RSP), because Anthropic originated the RSP form the concept page is built around. A regulated-by: line names the Bureau of Industry and Security in prose, since no page exists for it.
  • Google DeepMind — score 4 (in-degree 281). Thirteen typed links. The contradicts: [[concepts/training-data-walls]] edge is qualified in place as a tension the page's own evidence carries rather than a resolved disagreement.
  • Nvidia & TSMC — AI Compute Infrastructure — score 4 (in-degree 216). Twelve typed links, led by a nine-target deployed-by: line naming the developers and cloud operators the supply analysis sits underneath — the edge type the schema added for exactly this direction and which the page's untyped "Related concepts" list could not express. contradicts: [[concepts/ai-bubble-debate]] is qualified: the page's demand figures are cited on both sides.
  • Meta AI — score 4 (in-degree 193). Twelve typed links, including both instance-of: and contradicts: edges to Open-Source AI / Open-Weight Models, qualified in place — Meta is the largest-scale open-weights case and the clearest case of that commitment narrowing, so the page is evidence on both sides.
  • DeepSeek — score 4 (in-degree 100). Eleven typed links, including contradicts: [[concepts/export-controls-ai]] and instance-of: [[concepts/fast-follow-problem]].

Zero, and not re-audited. Wiki/_meta/lint-report.md reports 213 distinct broken targets, 0 referenced 2+ times, 213 referenced once. The candidate lane the skill directs the run to work is empty for the third consecutive run. The single-reference tail is the settled long tail of author surnames, institution shorthand and deliberate forward references. Sixth consecutive run reporting zero on this basis. The count moved 212 → 213 because pages created and edited since 2026-09-08 introduced one new single-reference forward target; none of the five pages touched this run introduced any (verified programmatically below).

Queued — foundational sources

Two primary documents, both verified against the issuing body's own domain and saved in full.

  • H.R. 9917, AI Kill Switch Act, 119th Congress — introduced text (2026-07-23) — score 8 (live +3, uncaptured primary document +3, US AI regulation core +2). The only cited primary-document URL across the last three days of digests with no Raw Sources/ file, reported by bin/primary-doc-scan.py in its precise lane, and named in the ## Primary documents section of the 2026-09-09 22:05 digest. Verified: congress.gov, HTTP 200, Dublin Core metadata self-corroborating (dc.number H.R.9917, dc.creator Rep. Lieu, Ted [D-CA-36], dc.coverage 2026-07-23), enrolling line and short title confirmed present. Saved in full: Raw Sources/AI Kill Switch Act, H.R. 9917, 119th Congress (Introduced in House, 2026-07-23).md. Queued: INGEST-ai-kill-switch-act-hr9917-2026.md.

Four corrections to AI Kill Switch Act (Lieu–Moran), which has stood since 2026-08-07 at sources_count: 9 built entirely from the sponsors' press release and news coverage. (1) The page says "Penalties run up to $20 million per violation." § 2220F(d)(2)(B) says "not more than $20,000,000 for each day on which such violation occurs", and that figure applies only to violations of the emergency-authority subsection; the general penalty at (d)(2)(A) is $2,000,000 per day. Both the structure and the periodicity are wrong, and "per violation" for "per day" is the precise degree-compression error class named in schema v4.7 — found here in a page built from a press release, which is what that rule predicts. (2) The revenue test is narrower and affiliate-aggregated than the page's "annual technology revenue", and is one of three conjunctive requirements. (3) The page presents "altering its safety rules" and "attempting to access its own weights" as statutory triggers; they are not prongs of "covered incident" but two of four illustrative limbs of the separately defined "loss-of-control scenario", and every prong is limited to occurrences "outside of red-teaming or other structured testing" — a carve-out the page omits entirely. (4) The page carries neither the bill number nor the referral. Ten further provisions present in the text and absent from the page are listed in the ingest task. The page also carries a malformed doubled (Source: (Source: ...)) citation in two places.

  • China MIIT, 《信息通信行业发展"十五五"规划》, 工信部规〔2026〕188号 — score 8 (live +3, primary document +3, compute core +2). This closes the item the 2026-09-09 report named "the most substantive miss" and deferred for one run because only a republication on a Beijing municipal site had been located. That report's instruction — "next run should target miit.gov.cn 政策文件 directly rather than re-searching generally" — was followed and worked: a site-scoped search returned the Ministry's own notice page on the first attempt. Verified: miit.gov.cn under its 政务公开 > 政策文件 > 文件发布 > 通知 path, HTTP 200, with the structured header self-corroborating every identifying field (发文字号 工信部规〔2026〕188号; 成文日期 2026-08-12; 发布日期 2026-09-07; 发布机构 信息通信发展司), and SiteIDCode bm07000001, MIIT's registered State Council site identifier. The plan body was pulled from the PDF attachment that same notice page embeds, so provenance is unbroken and same-origin: 20 of 20 pages, complete. Saved: Raw Sources/MIIT - Notice on Issuing the Information and Communications Industry Development 15th Five-Year Plan (Gong Xin Bu Gui 2026 No. 188, 2026-08-12).md, circular plus full plan, in the original Chinese. Queued: INGEST-miit-ict-15th-five-year-plan-2026.md.

The primary text confirms all three of the previous run's findings and adds a fourth. (1) It is the information and communications industry plan, not an AI plan — intelligent computing is indicator 9 of thirteen in 专栏 1, and the document's subject is networks, spectrum, submarine cables, satellites, emergency communications, telecom market regulation, anti-fraud and 6G at least as much as AI. (2) Two dates: 成文日期 2026-08-12, matching the signature block, against 发布日期 2026-09-07. (3) The 9,800 EFLOPS target has a denominator the digest omitted — 专栏 1 gives the 2025 baseline as 1,590 EFLOPS with a five-year cumulative change of 〔8,210〕, a rise of roughly 6.16×; the separately reported 2,185 EFLOPS at end-June 2026 is a different series and must not be mixed with it. (4) New, and not in any coverage the vault holds: 专栏 5 智能体互联网络建设 is a programme to build state identity and control infrastructure for AI agents — upgrading the DNS to establish an "internet agent identity identification system", blockchain and electronic certification for trusted authentication, "agent registration management, identity authentication, ownership determination, address resolution" capabilities, a "monitoring, early-warning and security-control capability over the full lifecycle of agents", and international nodes able to serve agents globally and interoperate with "mainstream international technical systems". No existing vault page has an equivalent. Also present and unheld: ten-thousand- and hundred-thousand-card cluster deployment with "greatly increased adaptation to domestically produced computing chips"; deep-synthesis forgery detection, training-data protection and agent security as named research priorities; an anti-fraud vertical large model; AI-assisted market-access approval; and network-native AI and token (词元) communication as 6G research priorities.

Authenticity-verification failures

None. Two documents sought, two verified and saved in full, none rejected.

One fetch failed for reasons unrelated to authenticity: parallelparliament.co.uk returned HTTP 403 behind bot verification when the September 8 Commons debate record was sought for Sobel's own words on introduction. Not a verification failure and not retried; the parliamentary record at bills.parliament.uk supplied the authoritative facts, and Hansard is the better target next time.

Deferred backlog (over the daily cap — re-surfaces next run)

New this run

  • entities/bernie-sanders (score 7). Named on 31 content pages — including Overview, Index, AI Industry Lobbying — The 2025-2026 Political Offensive, AI Political Cleavages, AI Public Wealth Fund and Government Equity in AI, AI Backlash and Ban Artificial Superintelligence Act (announced 2026) — with no page behind the name, and he is the central actor in two live threads this week: convening senators for an AI briefing, and the Ban Artificial Superintelligence Act. Alias resolution finds a genuine defect rather than a page under another name: two wikilinks read [[entities/sen-warren|Bernie Sanders]], pointing Sanders's name at Elizabeth Warren's page. That is a mis-targeted alias, not a broken link, so lint-scan.py cannot see it — it resolves cleanly. entities/sarah-huckabee-sanders also exists and is a different person. The right fix is a page plus repair of the two aliases; the wrong fix is repointing the aliases at a page that does not exist yet. Deferred on cap, not on merit, and it is the strongest single candidate for the next run.
  • Tech Transparency Project report on AI-generated CSAM advertising across Meta's platforms (score 6), and entities/tech-transparency-project (score 3) — second deferral. TTP is now named on two content pages (AI and Children, Meta AI), up from zero when first raised. The report remains the uncaptured instrument the September 8 litigation item is about.
  • Expanded class action against Anthropic over Max-tier usage limits (score 6) — second deferral, and the single second-lane hit from primary-doc-scan.py this run. Read rather than assumed: it is a genuine candidate, not an allusion. Still blocked on the same thing as last run — neither the court nor the case number is recorded in any digest, so a docket search must precede a pull, and the CourtListener MCP remains unauthenticated for the seventh consecutive run.
  • UK Artificial Superintelligence Bill, operative text as introduced (score 5). Not queued because it does not appear to exist publicly yet — bills.parliament.uk shows the bill at first reading with no publication located. Recorded as an open question on the page instead of as a queued capture. Also unqueued for the same reason: the Clement-Jones amendment to the Cyber Security and Resilience Bill (bills.parliament.uk/publications/67506/documents/8704, named in the 2026-09-09 digest's ## Primary documents list), and the Sanders–Casar press release, both of which are reachable and are the better next captures in this thread.

Carried forward

  • China MIIT plan — closed this run. Recorded here only to mark that a deferral reached action on its second run, which is what the deferred backlog is for.
  • legislation/ pages for the AI OVERWATCH Act, the Chip Security Act and the MATCH Act (score 5 each) — second deferral. Reachable only through FY27 National Defense Authorization Act (Chairman's Mark). H.R.6875 resolves on Congress.gov as the AI OVERWATCH Act; the Senate amendment numbers (S.Amdt. 6105, 6575, 6585) are in the ITI coalition letter captured 2026-09-09. The placement question — three pages or one section — is still undecided and is what the cap keeps deferring.
  • Global Dialogue on AI and the Independent International Scientific Panel on AI (score 4); entities/volker-turk now exists (created by the nightly fold since the last scan), so the associated entity gap has narrowed to an OHCHR page.
  • OpenAI Navier–Stokes proof artifacts (score 4) — second deferral. The proof write-up PDF, the Euler proof PDF and github.com/openai/NavierStokesAndEuler. The Alpöge–Buckmaster documents remain unlocated.
  • 葬AI / FuneralAI ChinaTalk essay (score 4) — second deferral.
  • entities/supreme-peoples-court and the SPC companion Q&A (score 3) — second deferral; the main SPC Opinion ingest is queued and China — SPC Opinions on Lawfully Adjudicating AI-Related Dispute Cases (法发〔2026〕10号) now exists.
  • Slice-0 thin anchors with in-degree ≥6 (score 2–3 each), new this run. Four companies/ pages under 350 words at meaningful in-degree: Coinbase (328 words, in-degree 10, confidence: high on 2 sources, unedited since 2026-06-06), Clearview AI (320, 7, medium/1), PrismML (256, 6, low/4). The fourth, Fairly Trained (303, 8, low/1), is excluded rather than deferred — it is the subject of a standing needs-review note from 2026-07-16 proposing migration to entities/, which is a curator decision, not a gap to re-raise.
  • Slice-0 health, for comparison against prior slices. Of 168 companies/ pages, only 5 lacked ## Relationships — and all five were worked this run, so the slice is now complete on that measure. This is by far the best-wired slice audited: against 15 of 374 in slice 11, 9 of 172 in slice 12 and 7 of 149 in slice 13. confidence: high on sources_count ≤ 1 returns just 2 of 168 here (Coinbase is high on 2; Stainless is high on 1), against 90 of 149 in slice 13 and 102 of 172 in slice 12. The pattern named in the last seven reports is a sources/ phenomenon, not a wiki-wide one, and slice 0 is the cleanest evidence of that yet — which strengthens rather than weakens the standing recommendation that a sources/-specific confidence rule belongs in the schema.
  • entities/adam-thierer (score 3) — thirteenth consecutive deferral, structural. Reported once, not re-listed.
  • Three slice-3 pages, twelfth deferral — structural: Standards-Industrial Complex, Three Theories of Victory (US / EU / China AI Governance), Transaction Costs (Coase) and the Agentic Economy. Twelve runs without reaching the top ten. They need a direct instruction or a decision that they are not gaps.
  • Slice carryover, unchanged. Slice-4 (tenth): American Civil Liberties Union (ACLU), Access Now. Slice-5 (ninth): Helen Toner, International Monetary Fund (IMF), Henry Farrell. Slice-6 (eighth): Stanford Center for Research on Foundation Models (CRFM), Robert Chesney, Sen. Elizabeth Warren. Slice-8: GEMA v. Suno (Munich Regional Court), Sony Music v. Udio (second suit, 2026) — both still need a docket check the unauthenticated CourtListener connector blocks. Slice-9: Grok Imagine (fifth deferral; recommended on 2026-09-08 for model-robustness-check rather than this skill, and that recommendation stands). Slice-10: Missing Coverage: Companies, People, Models, Legislation, Concepts (revised), Crystallize Session: Wiki Gaps & User Knowledge Map (2026-04-13).
  • Twenty-one source-robustness-check candidates across slices 11–13 — seventeenth consecutive report, structural. Out of this skill's lane.
  • Date-stamped headers in lint-report.md — twenty-fourth cycle, structural. Holding at 26. Self-references now 6.
  • Carried unchanged. entities/nightingale; legislation/caadca; entities/david-feith and entities/hudson-institute; Marvell; California AB 1856; the Texas OAG / Meta Webster–Mahoney exchange; the two chain-of-thought monitoring papers (arXiv:2507.11473, arXiv:2503.11926); the OpenAI June 2026 Threat Report PDF; the Epoch AI AI-R&D task taxonomy; the ARC Prize Astra evaluation post; Supreme Court of India, Vijay Ghanshyam Gadiya v. Union of India; litigation/worker-info-exchange-v-uber; entities/center-for-shared-ai-prosperity; CMORG "Firm Guidance For Frontier AI"; NCSC frontier-AI guidance; the Minnesota AG Emerging Technology report; ESRB/2026/3; the Governor of the Bank of England's 9 July 2026 letter to Dame Meg Hillier MP.

Verification notes

  • No broken outbound links introduced. All six touched content pages and both new queue files were resolved programmatically against the full page set after writing: zero unresolved introduced on every one. Six unresolved targets remain inside the bodies of the touched pages (concepts/anthropic-compute-strategy, entities/cosmos-institute, companies/aeon, concepts/ai-and-the-classroom, entities/dina-powell-mccormick, entities/lerrel-pinto) — all pre-existing members of the single-reference long tail, none touched. One unresolved target was caught in a queue file and fixed before it shipped: companies/huawei → Huawei — Ascend AI Accelerators.
  • Style scan clean. All six touched content pages checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each for banned headers, self-reference, date-led headers and predictions. Five hyperbole matches were flagged on Anthropic (3) and Google DeepMind (2); all five are pre-existing and all five sit inside quoted or attributed source material — "mind-blowing" inside a WSJ headline, "unprecedented" attributed to industry groups and to Goldman Sachs, "landmark" inside a publication's headline. House style governs the wiki's own voice, not quotations, so none was altered. None was introduced this run.
  • Revision fidelity. The five companies/ edits were purely additive apart from one last_updated value each. UK Artificial Superintelligence Bill (ControlAI draft, May 28, 2026; introduced September 8, 2026) is the only page whose prose changed: the status line and the ballot paragraph were rewritten because they asserted a status the instrument had already left behind, and every sourced fact in the original survives — the 20-MPs ballot, Swayne and Whittingdale, the contact tool, the PMB low-probability framing, the Fry statement in full, the comparison-to-other-regimes section and both original sources. Nothing was removed from any page.
  • Confidence set honestly, and not inflated anywhere. UK Artificial Superintelligence Bill (ControlAI draft, May 28, 2026; introduced September 8, 2026) moved 2 → 6 on sources_count but stays medium, because the introduced text is not held. No confidence or sources_count was changed on any of the five wired companies/ pages — graph wiring adds no evidence.
  • Source fidelity. Dates were taken from the documents' own faces, not from the digests: H.R. 9917 from congress.gov's dc.coverage and the enrolling line, the MIIT circular from its 成文日期 and signature block, the UK bill from Parliament's own passage record. Six divergences between primary text and existing wiki prose are recorded above and in each raw file's gap_scan.notes.
  • Capture completeness stated, not assumed. Both documents are saved in full. The MIIT capture is 20 of 20 PDF pages, in the Chinese original with no translation made; five 专栏 table cells clipped during PDF extraction are marked in place as *[cell clipped in source extraction]* rather than reconstructed. The H.R. 9917 capture is the complete introduced (ih) text; congress.gov states there is one version of the bill.
  • Cap respected. Eight gaps actioned against a 6–10 cap; 19 candidates found after dedupe. Six scrapes and three searches; no crawls; no failures other than the one 403 recorded above.
  • Lock held in file-only mode throughout, with heartbeats between phases. No git command was run from this session; the changeset is left uncommitted for the host sync loop.

One-line summary

UK Artificial Superintelligence Bill (ControlAI draft, May 28, 2026; introduced September 8, 2026) was recording a bill as "awaiting introduction" two days after it was introduced in the Commons, and is now correct, sourced to Parliament's own record, and carries the two-names and competing-firsts divergences unresolved; the MIIT five-year plan was pulled complete from the Ministry's own domain, closing last run's top deferral and confirming all three of its findings plus an unreported agent-identity-and-control programme; H.R. 9917's statutory text was captured and yields four corrections to a page built from a press release, including a $20 million "per violation" that the bill says is per day; and the five highest-in-degree companies/ pages — led by Anthropic at in-degree 819, the highest found in any slice — were wired into the typed graph, completing slice 0 on that measure. Needing the curator's attention: two wikilinks that point the name "Bernie Sanders" at Sen. Elizabeth Warren, which no broken-link scan can see because they resolve.

Categories: Briefings | briefing | gap-scan