Gap Scan — 2026-09-18

Scanned

Recent window: 4 New Developments Log/ files dated within 48 hours (2026-09-16 22:11 through 2026-09-17 22:05), and 137 wiki pages carrying last_updated: 2026-09-17 or 2026-09-18. Rotation slice: 8 — litigation/ + standards/ + government/, 97 pages.

Candidates after dedupe: 23. Actioned: 8. Web pulls: six Firecrawl calls (three searches, three scrapes, one of them a PDF parse) and four CourtListener calls; no crawls.

Gaps actioned (8 of 23 found)

New pages created (live)

• Protect Democracy Project v. Office of the National Cyber Director — score 7 (rotation slice +2 as a litigation/ gap in a wiki core area, live thread +3, backlog +1, in the deferred list since 2026-09-16). Gap type 3, third deferral, now closed. The 2026-09-16 report created Protect Democracy and recorded the FOIA action in prose only; the "Where does X go?" rules put an individual action in litigation/, and both the 2026-09-16 and 2026-09-17 reports deferred it on the cap while noting the CourtListener connector had made the docket retrievable.

Built from the docket rather than from coverage, which is what makes it worth more than a restatement of the entity page. Confirmed and added: the court (D.D.C.), the cause (5 U.S.C. § 552), nature of suit 895, no jury demand, counsel on both sides, the eleven complaint exhibits, and the reassignment from Judge Timothy J. Kelly to Judge Ana C. Reyes on September 3.

The substantive finding is the procedural turn nothing in the vault recorded: a preliminary-injunction motion was filed September 1 with six supporting declarations, a briefing schedule and a September 22 hearing were set on September 8–9, and then on September 15 the plaintiff withdrew the motion, the court vacated the hearing and schedule, and set a joint status report for November 13, 2026 — thanking the parties "for working to resolve this matter expeditiously." That wording, together with the September 8 production the organization reports, indicates a negotiated production rather than a setback, and it resolves how the September 30 date the complaint sought was disposed of: it was never ruled on. confidence: medium on 4 sources, six typed relationships, four ## Open questions.

• Andrew Ng — score 5 (live thread +3, safety/alignment core +2). Gap type 3. Zero mentions across all 1,800-plus content pages before this run — the clearest single absence found in the slice-independent sweep, for a figure the wiki's own AI Existential Risk page argues against by implication without naming.

Built from his own site and the Bloomberg report of his September 17 remarks: the Google Brain founding lead, the Baidu chief-scientist role and its 1,300-person team, the Stanford AI Lab directorship, Coursera, DeepLearning.AI, AI Fund, AI Aspire and LandingAI. The policy section carries the September 17 position — extinction warnings from frontier-lab researchers as "science fiction," and the claim that the industry stoked catastrophic-risk fears to bolster publicity and shape regulation — and says plainly that the last is an argument about motive rather than about the technical question, and that no response from those researchers is held. confidence: medium on 2 sources, and sources_count says 2 rather than being rounded up: the Bloomberg piece was read only to the lede.

One divergence is recorded rather than smoothed: his site's summary description identifies him as a board director at Amazon while its own biography text does not, and no other held source corroborates it. That is left as an open question, and it matters because Amazon issued its own AI safety position on September 16.

Pages expanded (live)

• Sony Music v. Udio (second suit, 2026) — score 4 (rotation slice, eighteenth-deferral carryover, confidence: low). The slice's weakest litigation page, and the one the last three reports named as unblocked by CourtListener without ever reaching the cap. The docket corrects or confirms almost every uncertain field on the page:

• Docket number established: No. 1:26-cv-06120, which the page did not have.
• Court confirmed: S.D.N.Y. The page had recorded it as "reported; venue follows the original S.D.N.Y. case."
• Filing date corrected: the complaint was filed July 20, 2026, not July 21. July 21 is the case-opening and ECF-designation date, which is what trade coverage reported. Both dates are now on the page with the distinction stated, rather than one silently replacing the other.
• Judge established: Alvin K. Hellerstein — and the page now records that the case was accepted as related to No. 1:24-cv-04777-AKH on July 24, so the judge who denied Sony's motion to expand the original case is the judge hearing the standalone suit. The page previously asserted the venue relationship as an inference; it is now a docket fact.
• Parties completed: ten Sony-affiliated plaintiff entities plus John Does, against one named plaintiff before; Sony Group Corporation identified as corporate parent in the Rule 7.1 statement.
• Service and deadline: served August 6, 2026, responsive pleading due August 27, 2026 — the page had said no responsive pleading "had been reported at the time of writing" with a last_status_check of July 22.

2,899 → 5,635 body chars; 1 new (Source: URL) cite; sources_count 3 → 4; confidence low → medium; last_status_check 2026-07-22 → 2026-09-18. Held at medium rather than raised further for a stated reason: docket entries 24 through 27, filed August 27 and August 31, appear in the public record without retrievable text, so whether Udio answered, moved to dismiss or obtained an extension is not established. The page says that instead of implying the docket ends at entry 23.

• Protect Democracy — expanded as a consequence of the new litigation page. One of its three ## Open questions is closed by the docket: the page said the court "has not been confirmed against the docket" and offered D.D.C. as a consistent guess; it is confirmed. The question about the September 30 production date is not closed but is now answered as far as the record allows — the motion that would have produced a ruling was withdrawn. The bare-text case reference became a [[litigation/protect-democracy-v-oncd]] wikilink in both the prose and the litigates: relationship. sources_count 4 → 5. No prose sentence, figure, name or existing citation was removed.

Link aliases fixed

Zero, and the lane consumed no budget. lint-scan.py reports 0 broken targets referenced 2+ times and 210 referenced once. This is the fifth consecutive scan at zero in the 2+ lane, and the settled finding stands: the single-reference tail is author surnames, institution shorthand and deliberate forward references. The tail was not re-audited.

Three alias resolutions were nonetheless done during writing, which is where they belong: companies/alibaba → Alibaba / Qwen Team, companies/deepseek → DeepSeek, concepts/deepfakes → Synthetic Media / Deepfakes. None was a page-level gap.

Queued — foundational sources

• Rep. Ro Khanna, letter to Alibaba Group Holding Limited on pacing the frontier (2026-09-16) — score 8, the run's highest. Gap types 1 and 9. The sole first-lane primary-doc-scan.py hit resolving to a genuine gap. Captured complete — 2 of 2 pages, full body, both footnotes and the signature block.

The document is the first the vault holds in which a Member of Congress asks foreign AI developers directly to join a pacing commitment, rather than asking the US government to negotiate one, and it argues explicitly against the position — tracked on several pages — that pacing cedes leadership to Beijing. It puts four numbered questions with a September 22, 2026 response deadline, and its footnote 1 resolves to the Amodei essay the vault already holds.

Provenance flagged, not smoothed. No copy was found on khanna.house.gov or any committee domain; two searches returned only the Washington Post's exclusive and its document repository copy. The Post's repository is a secondary host, in the same class as the DocumentCloud captures accepted before, and the raw file's frontmatter says so. Corroboration was done harder in consequence: two passages quoted verbatim by the Post were confirmed present in the fetched PDF.

Two capture caveats recorded in the raw file itself. (1) This is one letter of three — the Moonshot AI and DeepSeek copies were not published and are not assumed identical. (2) The letter is dated September 16; the digest records the letters as sent September 17, which is the Post's publication date. That is the send-date-as-event-date pattern source-fidelity exists to catch.

Saved: Raw Sources/Rep. Ro Khanna letter to Alibaba Group on pacing the frontier (2026-09-16).md. Queued: INGEST-khanna-letters-chinese-ai-labs-pacing-2026.md.

• Pew Research Center, Globally, More People Expect AI to Cause Job Loss Than Growth (2026-09-17) — score 6. Gap type 9. Verified on pewresearch.org, HTTP 200, live fetch forced; four named authors each resolving to a Pew staff page; published time 2026-09-17T09:59:56-04:00. Eleven separate figures in the digest were checked against the source and all match, including the 34-of-37 jobs finding, the 37%/41% medians, Israel as the sole primarily-excited country, the US 18–34 shift from 40% to 55%, the 31-point US ideological gap, and the full fieldwork construction.

Queued, not captured, under the large-document rule: the report is eight web chapters plus an appendix and a topline PDF, and a partial capture under a whole-report title would misstate what the vault holds. Chapter 1 was read in full for verification. The capture instruction names chapter 5, "Do people trust China, the U.S. or the EU to regulate AI?", as the priority after the report PDF — it is the chapter closest to the wiki's core subject and the digest does not summarise it at all. Verification record: Wiki/_meta/queue/gap-scan/proposed-sources/pew-global-views-of-ai-2026.md. Queued: INGEST-pew-global-views-of-ai-2026.md.

• Chesney & Citron, Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security — score 5, third deferral, now closed. Gap type 1, and the clearest dangling foundational reference in the vault: Liar's Dividend carries depends-on: … (**not in wiki**) — origin about its own origin text, at confidence: high. Verified on Boston University's institutional repository — Citron's own institution — with full bepress citation metadata, and cross-corroborated by the SSRN record the repository itself links.

The verification corrected a citation the vault has carried wrong. The authoritative record gives Robert Chesney & Danielle K. Citron, 107 California Law Review 1753 (2019), volume 107, issue 6, publication date 12-2019. The vault has it variously as "Citron & Chesney" and "2018/2019"; 2018 is the SSRN posting date, not publication, and Chesney is first-named. Queued, not captured, as a full-length law-review article. Queued: INGEST-chesney-citron-deep-fakes-2019.md.

Authenticity-verification failures

• Coalition letter of 19 advocacy groups and AI researchers opposing the Thune–Klobuchar framework (2026-09-16) — second rejection under protocol step 5. A search on the letter's distinctive quoted sentence returned exactly one result: the same Washington Post exclusive that carries its contents and does not link the document. No signatory organization's site surfaced a copy. The 2026-09-17 report rejected it on identical grounds and expected a signatory to post it within a day; two days on, none has. Nothing saved, nothing queued. This is an availability failure, not a finding of inauthenticity — the quotations are consistent and the signatory set is named. Recorded at Wiki/_meta/queue/gap-scan/needs-review/2026-09-18-coalition-letter-19-groups-second-rejection.md with a stopping rule: if a third run finds nothing, record once that the vault carries this position only through the Post's account, and stop re-scoring it.

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

New this run

• entities/danielle-citron, score 4. Co-author of the origin text of a high-confidence concept page, with no page of her own, while Robert Chesney has had one since before 2026-09-16. Surfaced by today's Deep Fakes verification and flagged in its record; held only because the run was at cap.
• companies/udio, score 4. Suno has a page; Udio — Uncharted Labs, Inc. — does not, despite being the defendant in Sony Music v. Udio (second suit, 2026) and in the 2024 RIAA-coordinated action, and despite litigation/isbell-v-suno, litigation/gema-v-suno, litigation/umg-v-suno and litigation/cox-v-sony all existing. A defendant company appearing across four litigation pages with no company page is the clearest asymmetry in slice 8.
• litigation/umg-v-udio, score 4. The original 2024 case, No. 1:24-cv-04777 (S.D.N.Y.), has no page, while today's expansion of its 2026 offshoot now cites it by docket number three times. The Suno half of the same June 2024 RIAA filing has UMG Recordings v. Suno (AI music training data). Verified retrievable through CourtListener in the course of today's work.
• entities/office-of-the-national-cyber-director, score 4. Lead defendant in the litigation page created today, the agency at the centre of AI Pre-Release Vetting, and the only one of the four defendant agencies with no page. Named in the new page's own open questions.
• Anthropic, Measurements for understanding the pace of AI development inside frontier labs (2026-09-17), score 6. Listed in the 09-17 digest's ## Primary documents and the source of its lead item — the R&D Automation Index, Claude "leading" 26% of Anthropic's own AI R&D up from under 1% in February, ~30,000 concurrent agents, 0.002% of over a billion decisions blocked. primary-doc-scan.py matched it to a held file; the match should be checked rather than trusted, given today's Pew finding that the possible-match lane errs in both directions. If it is genuinely absent it is a score-8 item next run.

Slice-8 structural findings

Slice 8 is the best-built slice in the rotation on every structural measure, and this is worth stating once rather than re-deriving:

• Zero thin anchors. No litigation/, standards/ or government/ page under 1,800 body characters sits at in-degree ≥6.
• Zero pages at in-degree ≥6 lacking ## Relationships — against 3 of 170 in slice 7 and 1 of 135 in slice 6. Graph wiring in this slice is complete.
• One page at sources_count ≤ 1 with in-degree ≥6 (SpaceXAI v. Bonta (AB 2013 challenge), in-degree 6, medium), and one page at confidence: high on sources_count ≤ 1 — against 22 and 23 respectively in slice 7. The single-source-high-confidence pattern that four consecutive slices reported as rising does not appear in slice 8, which is the first counter-example and worth recording as such.
• 11 of 97 pages carry last_updated: 2026-06-06 — the import signature identified in slices 3 through 7. Sixth consecutive slice, but the lowest proportion yet (11%, against 33% in slice 7 and 30% in slice 6). It remains one job, not eleven gaps.

Carried forward

• legislation/eo-trump-ai-cyber-2026-draft has no sources_count field, score 5 — second deferral. In-degree 73 at confidence: high. Still a schema defect on a load-bearing page that needs a citation audit, not a line edit. legislation/california-ccpa-regulations has the same defect.
• Letter of 107 House Democrats to Speaker Johnson (2026-09-16) — closed, not by this skill. Deferred at score 6 yesterday; INGEST-house-democrats-letter-johnson-ai-safeguards-2026.md is now in the queue, filed by developments-log with the letter captured complete including all 107 signature blocks. Recorded so the backlog stays honest. Third consecutive run in which the nightly fold closed a gap this skill deferred — the division of labour is working.
• DeepSeek-V4.1-Flash technical report (score 5) — fifth deferral. Still "possible match under another address." Today's Pew case shows that state is unreliable in both directions; this one needs a human look, and saying so a fifth time is not progress.
• Anthropic August 2026 Risk Report — recurring false positive, third statement. Flagged again by the possible-match lane against the 09-17 digest, which itself annotates the entry "a copy is held in Raw Sources/ under its canonical landing page." A scanner allowlist entry would close this permanently; it has now cost four reports a line each.
• Anthropic's Advanced AI Framework — new false positive in the precise lane. primary-doc-scan.py reported the CDN PDF path as having no raw file. Raw Sources/Anthropic's Advanced AI Framework.md and its .pdf have been held since 2026-06-21. Same URL-path mechanism as the Risk Report case, but in the precise lane rather than the possible-match lane, which is the first time that lane has produced one. The raw file's own frontmatter names INGEST-anthropic-advanced-ai-framework-2026, which is not in the queue — worth a curator look at whether that ingest was completed or lost.
• H.R. 10315 (Gottheimer, AI labs in K-12), score 6 — seventh deferral. Standing caveat unchanged: the digest cites an all-actions URL, not bill text.
• California SB 1119 and AB 1864 bill texts (score 6 each) — eighth deferral; the eleven other California child-safety bills signed September 10 (score 5 as a set) — eighth deferral, still a placement decision rather than research.
• Anthropic Institute economic scenarios working paper (score 5) — fifth deferral, a named hole in The Anthropic Institute's own open questions.
• entities/zerra-dc (score 3) — second deferral; entities/center-for-american-progress (4) and entities/bilal-chughtai (3) — third; entities/rum-group (4) — fifth; Chen Yixin / China Cyberspace (4) — fourth; South Korea's expanded espionage law (5) — fifth; Epoch AI, The Nvidia-sized hole in US GDP statistics (5) — sixth; 404 Media / Marshall Project DHS targeting teams (5) — fifth; Terence Tao, "A Severe Misalignment of AI in Mathematics" (5) — seventh; Nvidia–Groq DOJ antitrust investigation (5) — eighth; entities/bureau-of-land-management (4) — fifth.
• Tenth and eleventh deferrals: entities/tech-transparency-project and the TTP CSAM-advertising report (6); the expanded Anthropic Max-tier class action (6, unblocked by CourtListener and now three runs past that); legislation/ pages for the AI OVERWATCH Act, Chip Security Act and MATCH Act (5 each, placement undecided); the OpenAI Navier–Stokes proof artifacts; the 葬AI / FuneralAI ChinaTalk essay; entities/supreme-peoples-court and the SPC companion Q&A; the Clement-Jones amendment and the Sanders–Casar press release; OpenAI's Daybreak for Frontline Defenders and entities/ms-isac; entities/alex-sobel and the Canadian half of the ControlAI coalition; entities/cosmos-institute (fourth deferral).
• entities/adam-thierer (score 3) — twenty-first consecutive deferral, structural. Reported once, not re-listed.
• Three slice-3 pages, twentieth deferral — structural: Standards-Industrial Complex, Three Theories of Victory (US / EU / China AI Governance), Transaction Costs (Coase) and the Agentic Economy. source-robustness-check work.
• Slice carryover. Slice-4 (eighteenth): American Civil Liberties Union (ACLU), Access Now. Slice-5 (seventeenth): Helen Toner, International Monetary Fund (IMF), Henry Farrell. Slice-6 (sixteenth): Stanford Center for Research on Foundation Models (CRFM), Sen. Elizabeth Warren, plus Simon Willison's missing ## Relationships. Slice-7 (second): California CCPA/CPRA Regulations — Automated Decision-Making Technology (ADMT) and AI LEAD Act (S. 2937) missing ## Relationships. Slice-8: Sony Music v. Udio (second suit, 2026) closed today after eighteen deferrals; GEMA v. Suno (Munich Regional Court) carries forward — and it is now the older of the pair, at confidence: medium on 2 sources, though as a Munich Regional Court matter CourtListener does not reach it, which is why it did not close alongside its sibling. Slice-9: Grok Imagine (thirteenth; model-robustness-check work). 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 — twenty-fifth consecutive report, structural. Out of this skill's lane.
• Date-stamped headers in lint-report.md — thirty-second cycle, structural. Now 22, down from 25. Self-references 6, unchanged.

Findings that are not gaps

• The precise lane of primary-doc-scan.py produced its first false positive. Two hits against the 09-17 digest: the Khanna letters (genuine, actioned) and Anthropic's Advanced AI Framework (held since June 2026 under a different URL). Until now the precise lane had been reliable and only the possible-match lane produced these. The mechanism is the same — a CDN path cited where the vault holds the document under another address — so the lane is not broken, but it is no longer safe to work to zero without checking.
• The possible-match lane errs in both directions. The Anthropic Risk Report is a held document reported as possibly missing; the Pew survey was a missing document reported as possibly held, because three unrelated Pew files sit in Raw Sources/. Any publisher with several holdings will generate the second kind. The lane must be read, never trusted either way — and that is now demonstrated rather than asserted.
• entities/andrew-ng was a true zero, not an alias miss. Alias resolution was run before creating it: no page, and no mention of the name on any of the 1,800-plus content pages. For a wiki holding pages on considerably more peripheral figures, a complete absence rather than a thin page is the more surprising state, and it is recorded so the next run does not treat the new page as a duplicate of something.
• The single-source-high-confidence pattern broke its streak. Four consecutive slices reported it rising; slice 8 has one instance in 97 pages. The pattern is concentrated in entities/ and legislation/, not wiki-wide.

Verification notes

• No broken outbound links introduced. All four created or edited content pages were resolved programmatically against the full content-page set after writing: zero unresolved wikilinks on each. Three targets were corrected during writing rather than left broken (the alias resolutions above).
• Style scan clean. All four checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each, no exceptions to record.
• Revision fidelity. Three pages were edited rather than created. On Sony Music v. Udio (second suit, 2026), Protect Democracy and the queue files, no prose sentence, figure, name, quotation, table row, citation or existing relationship line was removed. Every (Source: …) present before the run survives: Sony v. Udio 3 → 4, Protect Democracy 4 → 5. Two claims were changed rather than only added — the Sony filing date (July 21 → July 20) and the Protect Democracy court (unconfirmed → D.D.C.) — and in both cases the prior reading is preserved in place with the reason for the divergence, so the change is auditable.
• Confidence raised once, deliberately. Sony Music v. Udio (second suit, 2026) rose low → medium on a docket that established its court, judge, docket number, parties and service history. It was not raised to high, and the report says why: four docket entries have no retrievable text, so the current status is genuinely unknown. Two pages created at medium, never high.
• sources_count set by count, not by feel. Protect Democracy v. ONCD 4, Andrew Ng 2, Sony v. Udio 4, Protect Democracy 5 — each the number of distinct cited URLs on the page, checked programmatically after writing. The Andrew Ng page was corrected downward from 3 to 2 by that check rather than left flattering.
• Capture completeness stated, not assumed. Khanna letter: complete, 2 of 2 pages, with the one-of-three caveat on the face of the file. Pew: not captured, queued with three verified URLs and a named capture priority. Chesney & Citron: not captured, queued with the repository and SSRN URLs.
• Cap respected. Eight gaps actioned against a 6–10 cap; 23 candidates after dedupe. Ten tool-backed retrievals in total; no crawls. The broken-link lane consumed no budget, as intended.
• 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

Two pages created and two expanded live, three foundational sources verified and queued (one captured complete), one source rejected a second time for availability — and two scanner findings worth more than any single page: the precise lane of primary-doc-scan.py produced its first false positive, and the possible-match lane is now demonstrated to err in both directions, which needs the curator's attention before the next run works either lane to zero.