AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-09

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

Contents

Scanned

Recent window: 3 New Developments Log/ files in the last 48 hours (2026-09-07 22:05, 2026-09-08 08:09, 2026-09-08 22:05). Rotation slice: 13 — sources/ Q–Z plus the Raw Sources/ orphan check (149 pages).

Mechanical inputs: bin/lint-scan.py (212 distinct broken wikilinks, 0 referenced 2+ times, 212 referenced once — the referenced-2+ candidate lane is empty for the fifth consecutive run, so no broken-link candidates were generated and the single-reference long tail was not re-audited, per the step-3 exclusion); bin/primary-doc-scan.py --days 3 (6 digests, 5 cited primary-document URLs, 0 missing from Raw Sources/, 1 possible match under another address, 2 lane-2 keyword pre-filter hits); bin/ingest-backlog.py (895 raws complete, 0 foundational without a sources/ page, 0 unclassified, 1 non-markdown raw unconverted — the Leo XIV encyclical, a known structural item).

The Raw Sources/ orphan check, which slice 13 owns, returns zero. Every foundational raw has a sources/ page or an open ingest task. The five INGEST- tasks queued on 2026-09-07 and the three on 2026-09-08 have all been processed; one task (INGEST-brinkema-adtech-remedy-memorandum-2026.md) remains open by design, held until September 16 when the sealed memorandum's redaction process runs.

22 candidates found after dedupe; 8 actioned.

Gaps actioned (8 of 22 found)

New pages created (live)

Pages expanded (live)

Three slice-13 sources/ pages at high in-degree with no ## Relationships section. All three edits are 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 is not new evidence.

  • Stanford HAI AI Index Report 2026 — score 2 (in-degree 139, the highest in the slice lacking the section). Eleven typed links. Two carry a claim rather than a bare pointer: supports: [[concepts/fast-follow-problem]], because the report's 2.7% March 2026 figure is the quantification of what that page argues qualitatively; and contradicts: [[sources/ai-as-normal-technology]], qualified in place because the report cuts both ways — 53% population-level adoption in three years runs against the slow-diffusion thesis while the "jagged frontier" of uneven capability runs with it.
  • We Need a Science of Scheming — score 4 (in-degree 73, safety/alignment core +2). Ten typed links tracing the empirical chain the article called for: depends-on the earlier Frontier Models are Capable of In-context Scheming, supports the later Stress Testing Deliberative Alignment for Anti-Scheming Training, and related edges to the Anthropic sabotage report and METR's review, which supply the evaluation-versus-deployment gap the article's oversight-evasion pressure predicts.
  • Raine v. OpenAI — Wrongful Death Complaint (2025) — score 2 (in-degree 43). Nine typed links, and a style repair: the page carried a ## Relevance to wiki header, which is banned self-reference under house style §2d. Its five items were converted into typed relationships with every fact and the Anthropic wellbeing citation preserved, and two edges were added that the old section lacked — Raine v. OpenAI, Inc., the case-tracking page for the same action, and Garcia v. Character Technologies, Inc. alongside the source page for the parallel Character Technologies suit.

Zero, and not re-audited. The referenced-2+ lane of Wiki/_meta/lint-report.md is empty; the 212 single-reference targets are the settled long tail of author surnames, institution shorthand and deliberate forward references. Fifth consecutive run reporting zero on this basis.

Queued — foundational sources

Four primary documents behind live-thread digest items, all verified against the publishing organization's own domain and saved in full.

  • OpenAI, "On the Navier–Stokes Millennium Prize Problem" (2026-09-08) — score 8 (live +3, uncaptured primary document +3, frontier-model core +2). The lead item of the 2026-09-08 22:05 digest and its 📌 update both rested entirely on this post plus Wired and Axios coverage; nothing first-party was held. Verified: openai.com, HTTP 200, dateline and og-metadata self-corroborating, distinctive passages confirmed present. Saved: Raw Sources/OpenAI - On the Navier-Stokes Millennium Prize Problem (2026-09-08).md. Queued: INGEST-openai-navier-stokes-2026.md. The post carries figures the digest condensed — 4.9 million agent messages and ~300 billion output tokens across all attempted problems, 2.7 million and ~130 billion for Navier–Stokes alone, and a separate earlier Euler result from nearly 100 agents over ~50 hours. A conflict on the face of the sources is preserved rather than resolved: the post calls the Alpöge–Buckmaster result "a resolution of the forced Euler problem" while Bubeck is quoted in Wired referring to "their work on unforced Euler," and the post separately says OpenAI's own agents resolved the unforced version.
  • Supreme People's Court of China, 《关于依法审理涉人工智能纠纷案件的意见》, 法发〔2026〕10号 (2026-09-07) — score 8 (live +3, primary document +3, AI-regulation core +2). The vault held no Chinese judicial AI material of any kind: a search across all content folders for "Supreme People" returned zero pages. Verified: court.gov.cn, which its own metadata describes as the SPC's only official identity on the internet; the page carries its source line (最高人民法院新闻局), timestamp (2026-09-07 19:22:29), issuing number and signature block, and the 24-article count is confirmed internally against the drafting note. Saved in full, in the original Chinese, with the News Bureau drafting note: Raw Sources/Supreme People's Court of China - Opinions on Lawfully Adjudicating AI-Related Dispute Cases (Fa Fa [2026] No. 10, 2026-09-07).md. Queued: INGEST-spc-ai-disputes-opinion-2026.md. The document is more specific than the coverage on training data. The digest recorded that the guidance "did not rule on ... whether training on copyrighted works without permission constitutes infringement"; article 6 does address training data on the personal-information side, holding that processing lawfully public personal information within a reasonable scope, absent express refusal, is generally not an infringement, with consent required where the effect on individual rights is major. The two genuine omissions — AI-output copyrightability and copyright training-data infringement — are framed by the SPC itself as deliberate 留白, blank-leaving pending accumulated experience, which is a stronger claim than "did not rule on."
  • ITI, CCIA, CTA, TIA and USCIB, letter to the Armed Services Committees on FY2027 NDAA export-control amendments (2026-09-03) — score 8 (live +3, primary document +3, compute/export-controls core +2). Verified: itic.org, HTTP 200, application/pdf, 3 pages, dated on its own face and naming its addressees, five signatories and ten cc recipients. Saved in full: Raw Sources/ITI, CCIA, CTA, TIA and USCIB - Letter to the Armed Services Committees opposing NDAA export-control amendments (2026-09-03).md. Queued: INGEST-iti-coalition-ndaa-export-controls-letter-2026.md. Two corrections to what the digest carries. (1) The digest and the Axios report attribute the letter to ITI alone; it is a joint letter over five signatures. (2) The digest records that the AI OVERWATCH Act's "congressional review period ... has been shortened from two years to 18 months, which ITI described as 'an eternity'." The letter says the Act "would set U.S. semiconductor controls in law for eighteen months" and calls eighteen months "an eternity"; it does not state the period was shortened, and does not use the phrase "congressional review period." Axios sources the shortening separately, to a Senate page — so the claim is supported, but not by the instrument, and a future page should not cite the letter for it. The letter also supplies the three amendment numbers (S.Amdt. 6105, 6575, 6585) the coverage omits.
  • Volker Türk, Global Update to the 63rd session of the Human Rights Council (2026-09-07) — score 8 (live +3, primary document +3, global AI governance core +2). Verified: ohchr.org, HTTP 200, speaker, forum, location and date stated on the page, quoted sentences confirmed present, and a trilingual PDF published alongside. Saved in full: Raw Sources/Volker Turk - Global Update to the 63rd session of the Human Rights Council (OHCHR, 2026-09-07).md. Queued: INGEST-turk-hrc63-global-update-2026.md. One misplacement corrected and one commitment recovered. The digest folds the call for prohibition of "weapons that can take lives without human involvement" into the AI item; in the speech it sits in the Ukraine section, prompted by reports of fully autonomous Russian drones killing three Ukrainians in August 2026, and Türk notes that "Ukraine has reportedly also field-tested such weapons" — both the theatre and the attribution to both parties were lost. Recovered from the primary text: "I will be writing to AI companies in the coming days," a dated and checkable undertaking; and the framing "I share the concerns of industry insiders that advanced AI could pose an existential risk," which locates the view as adopted rather than independently reached.

Authenticity-verification failures

None. Four documents sought, four verified and saved. One document was identified but not pulled — see the MIIT item in the deferred backlog; that is a host-preference deferral under protocol step 2, not a verification failure.

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

New this run

  • China MIIT, 《信息通信行业发展"十五五"规划》 (score 8) — the most substantive miss, and the identification work is done. The Compute item of the 2026-09-08 22:05 digest is about this plan. Two findings before it is pulled. (1) The instrument is not what the digest calls it. The digest says MIIT "released a five-year industry plan"; it is the information and communications industry five-year plan, in which intelligent computing is one of thirteen headline indicators — not an AI plan. (2) It is dated 2026年8月12日 and released publicly on September 7, so the digest's release date is right but the instrument's own date is three and a half weeks earlier. The digest also omits the 2025 baseline of 1,590 EFLOPS, without which the 9,800 EFLOPS target has no denominator; the plan sets a rise of roughly 6.2× over five years, and the separately reported 2,185 EFLOPS at end-June 2026 is a different series. Not pulled because the only copy located is a republication of the MIIT circular on ncsti.gov.cn, a Beijing municipal science-innovation site; protocol step 2 prefers the primary host, and miit.gov.cn's own policy-document page for this circular was not located in two searches. Next run should target miit.gov.cn 政策文件 directly rather than re-searching generally.
  • Expanded class action against Anthropic over Max-tier usage limits (score 6), a lane-2 pre-filter hit. Filed September 8, 2026 by Monica Vaca and Kati Daffan, both formerly of the FTC, following a July complaint withdrawn and refiled. The digest carries the complaint's allegations and a quotation from Anthropic's motion to dismiss in the earlier case; neither document is held, and the court and case number are not recorded anywhere in the digest, so a docket pull needs a search first. The vault has no litigation/ page for it.
  • Tech Transparency Project report on AI-generated CSAM advertising across Meta's platforms (score 6), and entities/tech-transparency-project (score 3). The report is the instrument the September 8 litigation item is about — more than 250 further advertisements since the start of August, a documented total above 350 since end-2025, around 300 directing users to AI image or video generation apps mostly linked to China-based developers. TTP is named on zero content pages, so it is a genuine entity gap as well as an uncaptured document. Deferred on cap, not on merit.
  • legislation/ pages for the AI OVERWATCH Act, the Chip Security Act and the MATCH Act (score 5 each). All three are reachable only through FY27 National Defense Authorization Act (Chairman's Mark), and are now the subject of a public split between a frontier lab and its trade association. H.R.6875 resolves on Congress.gov as the AI OVERWATCH Act, introduced 2025-12-18; the Senate amendment numbers are recorded in the captured letter. A page apiece is probably right, but the choice between three pages and one section on FY27 National Defense Authorization Act (Chairman's Mark) is a placement call the cap did not allow this run.
  • Global Dialogue on AI and the Independent International Scientific Panel on AI (score 4). Both welcomed by name in the Türk statement as recent international progress; neither is reachable in the vault. entities/volker-turk (score 3) and an OHCHR page are the associated entity gaps.
  • OpenAI Navier–Stokes proof artifacts (score 4). The proof write-up PDF, the Euler proof PDF and the Lean formalization repository at github.com/openai/NavierStokesAndEuler are the mathematical content behind the captured post and are not held. The Alpöge–Buckmaster documents posted 2026-09-07 are unheld and unlocated; a search for them is the prerequisite.
  • 葬AI / FuneralAI ChinaTalk essay (score 4), the second lane-2 pre-filter hit. An argument that open weights have changed the strategic position of Chinese model developers, cross-posted to ChinaTalk on September 6. Read and judged a genuine candidate rather than an allusion, but below the cap.
  • entities/supreme-peoples-court and the SPC companion Q&A page (score 3). The Q&A at court.gov.cn/zixun/xiangqing/511111.html carries the officials' own explanations, including the 避风港 (safe-harbour) framing of article 7. Deferred with the main ingest queued.

Carried forward

Verification notes

  • No broken outbound links introduced. All four touched content pages were resolved programmatically against the full page set after writing: zero unresolved on every one. The four new queue files carry seven hedged forward references (concepts/multi-agent-systems, legislation/ab-2013-training-data, concepts/open-weights, concepts/deepfake-regulation, entities/united-nations, concepts/international-ai-governance), each phrased in the task body as a target to decide on rather than an existing page. Queue files are operational, not content, and are excluded from the broken-link count.
  • Style scan clean. All four touched content pages checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each, including after the removal of the ## Relevance to wiki header.
  • Revision fidelity. The three expanded pages were edited additively apart from one last_updated value each and the single header conversion on Raine v. OpenAI — Wrongful Death Complaint (2025), where all five original items and the Anthropic wellbeing citation survive in the new typed section. No prose, figure, table row or citation was removed from any page.
  • Confidence set honestly, and not inflated anywhere. The new page is medium on 7 sources, not high — a fresh page does not meet the independent-3-source bar. None of the three expanded pages had confidence or sources_count changed, because graph wiring adds no evidence.
  • Source fidelity. Dates were taken from the documents' own faces, not from the digests: the ITI letter from its dateline, the SPC Opinion from its signature block and issuing number, the Türk statement from the OHCHR page's own date field, the OpenAI post from its dateline. Four divergences between primary text and digest are recorded above under the queued items and in each raw file's gap_scan.notes.
  • Capture completeness stated, not assumed. All four documents are saved in full text. Figures in the OpenAI post are reduced to bracketed captions; the SPC capture is the Chinese original with no translation made; the Türk capture is the English version, not the trilingual PDF. Nothing was summarised or trimmed in any of the four.
  • Cap respected. Eight gaps actioned against a 6–10 cap; 22 candidates found after dedupe. Seven scrapes and four searches; no crawls; no failures.
  • 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

Four primary documents behind live-thread items — OpenAI's Navier–Stokes post, China's first Supreme People's Court AI adjudication rules, the five-association NDAA export-controls letter, and Türk's HRC63 global update — were verified, captured in full and queued for review; ITI — Information Technology Industry Council was created live and three high-in-degree slice-13 sources/ pages were wired into the graph; the China MIIT five-year ICT plan is identified and deferred pending a primary-host copy, and four source-fidelity divergences between the digests and the underlying documents are recorded for the ingests to correct.

Categories: Briefings | briefing | gap-scan