Gap Scan — 2026-09-17

Scanned

Recent window: 4 New Developments Log/ files (2026-09-15 08:05 and 22:05; 2026-09-16 08:05 and 22:11), 130 wiki pages carrying last_updated within the last 48 hours. Rotation slice: 7 — legislation/ (all), 170 pages.

Tooling: bin/lint-scan.py (authoritative for broken wikilinks) and bin/primary-doc-scan.py --days 3 were both run rather than re-derived. Eleven web pulls: three Firecrawl searches, seven scrapes, one PDF parse. No crawls.

Gaps actioned (8 of 25 found)

New pages created (live)

• Ratepayer Protection Act (H.R. 9340) — score 8 (live thread +3, uncaptured primary document +3, US AI regulation core +2). Gap types 3, 8 and 9 together. The Act passed the House 417–3 on 2026-09-16 (Roll no. 312) and was already carried in prose on three content pages — AI Data Centers, Energy and Electric Power Sector and Mike Johnson — entirely through news coverage, with no legislation/ page behind it. Built from the reported bill text rather than from the coverage: the new PURPA § 111(d)(22) in full, both operative subparagraphs, the definition's two-limb test, and the § 112(b)(9) and § 112(i) conforming amendments. confidence: medium on 10 sources, 14 cites, a > **Provenance note:** recording that citations upgrade once the queued ingest runs, four ## Open questions, five typed relationships.

Two corrections the bill text makes to the coverage, and they run the same way — the instrument is weaker and broader than the item describes.

(1) The mechanism is PURPA consider-and-determine, not adoption. The digest says the bill establishes "standards state regulators could adopt to require AI data centers … to cover the costs." Section 2(b)(1) obliges each State regulatory authority and nonregulated utility to commence consideration within one year and to complete the consideration and make the determination within two. A State may consider the standard and decline it; the federal duty is to hold the proceeding. New § 112(i) then exempts a utility outright where the State has already implemented a comparable standard, already held a proceeding on one, or where the legislature has already voted on one — so an unknown share of the country is outside even the procedural requirement on day one.

(2) The covered class is not defined by AI. Subparagraph (C) defines a "large-load customer" by function and size: a non-residential consumer taking power "primarily to operate information technology infrastructure and related systems pertaining to data storage and computational applications and services," at 100 MW or more aggregate peak demand at a single site or campus. The operative text contains neither "artificial intelligence" nor "data center." "Data centers" appears only as a parenthetical example in the CRS summary. The standard reaches a 100 MW non-AI facility and misses a sub-100 MW AI one.

Four further divergences are recorded in the raw file: sponsorship (Evans is sole sponsor, Castor the original cosponsor, with 40 Members added 2026-09-10 — the digest gives both as introducers); the stranded-cost limb in subparagraph (A), which the digest does not carry; the codification targets, which it does not name; and the asserted relationship to the White House "Ratepayer Protection Pledge," which appears in coverage and nowhere in the bill.

• Future of Artificial Intelligence Innovation Act of 2026 (S. 3952) — score 5 (live thread +3, US AI regulation core +2), second deferral, now closed. The 2026-09-16 report left this as the unclosed half of a pair and gave a specific instruction: test S. 3952 against congress.gov directly rather than searching generally, as had worked for S. 5105 the day before. The test succeeded on the first call. S. 3952 is the Future of Artificial Intelligence Innovation Act of 2026, introduced 2026-02-26 by Sen. Todd Young "for himself, Ms. Cantwell, Mrs. Blackburn, and Mr. Hickenlooper," referred to Senate Commerce, star print ordered 2026-03-19, status still Introduced.

Two things the identifier settles that the coverage did not. The measure has four sponsors, not two — Blackburn and Hickenlooper were not named anywhere in the vault. And it predates the September negotiation by seven months, which places it as a standing vehicle the negotiation is reaching for rather than a product of it.

The page is built from the long title, the three-title structure, the fifteen section headings as printed in the bill, and a verified extract of section 101. confidence: medium on 2 sources, a provenance note, four ## Open questions, eight typed relationships. Section 101 is the substantive core: it places the Center for Artificial Intelligence Standards and Innovation within NIST, enumerates six functions, and carries a construction clause providing that nothing in it gives the Director "any enforcement authority that was not in effect the day before" enactment. That clause, under a title headed "Voluntary Artificial Intelligence Standards," is the element that bears directly on the September dispute over who runs frontier-model testing — and it cuts against reading the bill as a vehicle for mandatory government evaluation.

• DeepMind Institute — score 5 (live thread +3, frontier-lab governance core +2). Gap type 3. Launched 2026-09-16 and absent from the wiki entirely — zero mentions across all content pages. Built from two independent sources rather than one: the Axios launch account and the Institute's own site at institute.deepmind.com, which supplies the five inaugural essay titles the coverage does not give. confidence: medium on 4 sources, three ## Open questions, six typed relationships including one to The Anthropic Institute as the comparable body at another developer.

One mapping deliberately not asserted. Axios describes the inaugural publications as covering "economic policy for AGI, model reasoning transparency, global access, and human flourishing" — four topics. The site carries five essays, two of which are titled Principles for a new utopianism and A framework for frontier AI and the dawning of a new age. Both lists are recorded; the correspondence between Axios's last two descriptions and those two titles is plausible and is stated by neither source, so the page says so rather than merging them.

Pages expanded (live)

• California AB 3030 (Healthcare AI Disclosure) — score 3 (in-degree 18 +2, backlog +1). Graph wiring, not content. The page is the highest-in-degree page in the 170-page slice lacking a ## Relationships section, and the section it had was headed ## Related with five untyped related: lines. Renamed to ## Relationships, all five existing lines preserved verbatim, and two typed edges added: instance-of: [[concepts/ai-transparency]] and related: [[industries/healthcare]]. last_updated bumped to 2026-09-17 per the house-style checklist. No claim on the page was re-verified and no fact or citation was added or removed — this is a wiring edit, and confidence: high / sources_count: 6 were left untouched so the bump does not read as reinforcement.

Link aliases fixed

Zero, without re-auditing the list. bin/lint-scan.py reports 0 broken targets referenced 2+ times against 211 distinct broken targets, all single-reference long tail — which step 3 places permanently out of scope. This is the ninth consecutive scan to report zero on the settled 2026-08-18 basis. The distinct count moved 212 → 211 since yesterday, consistent with the nightly fold resolving one reference; no candidate entered the 2+ lane.

Queued — foundational sources

Four documents. Three captured complete; one deliberately not captured, for the reason given. Verification records for all four are in Wiki/_meta/queue/gap-scan/proposed-sources/.

• OpenAI, Our framework for reporting model misalignment (2026-09-16) — score 8, the run's highest. Gap types 1 and 9. The first published attempt by a frontier lab at a standing misalignment-disclosure procedure, named in the 22:11 digest through two sources, and absent from the wiki entirely. Verified: openai.com, OpenAI's own domain; the six incident reports resolve to the alignment.openai.com subdomain under per-incident slugs matching their summaries; independently reported the same day by the Wall Street Journal. Saved complete: Raw Sources/OpenAI - Our framework for reporting model misalignment (2026-09-16).md. Queued: INGEST-openai-misalignment-reporting-framework-2026.md.

Capture is partial by design and the raw file says so at its head. The framework post is complete; the six underlying reports are not captured, and the ingest task asks whether each is separately foundational before folding — three of them carry figures and a named model (27 affected summaries; GPT-5.6 Sol) that pages would cite by name.

The document carries a limit that must travel with it. OpenAI states the six "are reports of individual instances, and shouldn't be considered reflective of how often misalignment occurs across our models," and that they are "not intended to represent the full range or severity" of covered cases. It also does not describe them as previously unreported, as the digest does; the framework expressly contemplates re-disclosing behavior already carried in system cards. Separately, the post contains a company position on pacing — that OpenAI does "not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer" — published four days after the von der Leyen remarks and directly comparable to We Must Pace the Frontier (Amodei, September 2026), which the vault holds and which currently has no OpenAI counterpart.

• Gallup and SCSP, American AI Sentiment in 2026 (2026-09-15) — score 6. Gap type 9, and the only precise-lane hit from primary-doc-scan.py --days 3: one cited primary-document URL across six digests with no Raw Sources/ file. Verified: scsp.ai, SCSP's own domain, HTTP 200, application/pdf, 12 pages; self-corroborating through the Gallup copyright page, the Gallup colophon, and a methodology section giving the panel basis. Saved complete, all 12 pages including every chart's underlying data table: Raw Sources/Gallup and Special Competitive Studies Project - American AI Sentiment in 2026 (2026-09-15).md. Queued: INGEST-scsp-gallup-american-ai-sentiment-2026.md.

Every figure in the digest item checks out; one characterization does not. Use 39→51, knowledge 60→68, growth 36/45 reversing 53/24, jobs 70/9, businesses 59/14 — all confirmed. But the digest says trust "fell to 28% from 32%," and the report declines to say that: its words are that trust "has not increased," that 28% is "broadly in line with" 32%, and that trust is "stable and low." The gap exceeds the ±1.9-point margin of error, so the direction is not unsupported — it is the digest's inference, not Gallup's finding, and should not be attributed to Gallup. The sample is also more precise and more robust than the digest states: 3,152 adults, fielded June 2–23, 2026, on a probability-based panel the report expressly says is "not an opt-in panel."

The report's policy findings bear directly on a live wiki thread: 75% say safety testing should be conducted by independent experts, against 52% for government agencies and 35% for company self-testing, and 79% say government should prioritize safety rules even if that slows development. That is the public-opinion counterpart to the Senate testing dispute tracked at AI Pre-Release Vetting, Maria Cantwell and John Thune.

• H.R. 9340, Ratepayer Protection Act, text as reported (2026-09-10) — score 8. Verified: congress.gov, HTTP 200, Dublin Core metadata self-corroborating (dc.number H.R.9340, dc.creator "Rep. Evans, Gabe [R-CO-8]", dc.coverage 2026-06-18), the printed text's own Union Calendar No. 713 and [Report No. 119–814] matching the action history, the CRS summary restating the operative provisions subparagraph for subparagraph, and all three codification targets resolving to the U.S. Code. Captured complete — short title, section 2(a) subparagraphs (A)–(C), section 2(b) conforming amendments. Queued: INGEST-ratepayer-protection-act-hr9340-2026.md.

Version caveat recorded at capture. No engrossed (EH) print was posted; congress.gov listed only the reported and introduced versions, and a request for /text/eh returns the reported text. The captured version is the text the House passed "as amended," which is the right one, but the ingest is told to diff an engrossed print against it if one appears.

• S. 3952, Future of Artificial Intelligence Innovation Act of 2026 — score 5. Verified: congress.gov, both landing and text pages, HTTP 200, Dublin Core self-corroborating, sponsorship line confirmed inside the bill text itself. Queued rather than captured, under the large-document rule: the text runs to roughly 79,000 characters across three titles. Verification record: Wiki/_meta/queue/gap-scan/proposed-sources/future-of-ai-innovation-act-s3952-2026.md. Queued: INGEST-future-of-ai-innovation-act-s3952-2026.md, with section 101 named as the first capture priority.

One thing the CRS apparatus does not supply here: the summary is listed as "A summary is in progress," so no CRS gloss exists to corroborate the text against. That is why the page was written from the bill's own long title and printed section headings rather than from a summary, and why the ingest task is explicit that sections 102 and 103 are separate provisions with adjacent subject matter and must not be merged.

Authenticity-verification failures

• Coalition letter of 19 advocacy groups and AI researchers to Senate leadership opposing the Thune–Klobuchar framework (2026-09-16). Sought as a type-1/type-9 primary document: the 22:11 digest quotes it directly ("We strongly request that any AI regulatory framework empower the government, not AI companies…"; leading AI companies "cannot be trusted with setting the rules that protect public safety") and names six signatories including Encode AI, the Alliance for Secure AI, Issue One, the Institute for Family Studies, Alex Bores and Daniel Kokotajlo of the AI Futures Project. Rejected under protocol step 5: no copy on any trustworthy host was located. Two searches — one on the quoted sentence, one on the signatory set — returned exactly one result each, in both cases the same Washington Post "Exclusive" that carries the letter's contents and does not link the document. No signatory organization's own site surfaced a copy. Nothing was saved and nothing was queued; this is an availability failure rather than a finding of inauthenticity, and it re-enters scoring next run, when a signatory may have posted it.

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

New this run

• Letter of 100+ House Democrats to Speaker Mike Johnson (2026-09-16), score 6. Gap type 9. Led by Reps. Trahan, Liccardo, Whitesides and Lieu, urging the House stay in session until "meaningful, bipartisan AI safeguards" pass, and reaching kill-switch requirements and an antitrust waiver for safety collaboration — the same instrument family as the S. 5105 page built on 2026-09-16. Carried in the digest through Politico only, with no primary URL cited. Deferred on the cap; likelier than the coalition letter to be posted by a sponsoring office.
• legislation/eo-trump-ai-cyber-2026-draft has no sources_count field at all, score 5. The slice's highest-in-degree page at 73 inbound links, and among the highest in the wiki, carrying confidence: high with no source count in frontmatter to support it. This is a schema defect on a load-bearing page, not a thin-page problem — the body is substantial. Deferred because setting the field honestly means auditing the page's citations, which is a job rather than a line. legislation/california-ccpa-regulations (in-degree 9) has the same defect.
• Two further slice-7 pages at in-degree ≥6 with no ## Relationships, score 2 as a set: California CCPA/CPRA Regulations — Automated Decision-Making Technology (ADMT) (in-degree 9) and AI LEAD Act (S. 2937) (in-degree 6). Three of 170 pages in the slice, against 1 of 80 in slice 6 — graph wiring in legislation/ is close to complete.
• entities/zerra-dc, score 3. The Singapore-based developer of the 2.16 GW Australian campus Anthropic leased, named in the 2026-09-16 22:11 digest. One source, two people familiar with the deal, subject to Foreign Investment Review Board approval. Held at the quality gate.

Slice-7 structural findings

• Zero thin anchors. No legislation/ page under 1,800 body characters sits at in-degree ≥6. The folder is the best-built slice in the rotation on that measure.
• 22 pages at sources_count ≤ 1 with in-degree ≥6, and 23 pages at confidence: high on sources_count ≤ 1 (score 2 each as sets). The single-source-high-confidence pattern outside sources/ is now confirmed in four consecutive slices and is at its highest absolute count yet. The heaviest instances are Executive Order 14179 — Removing Barriers to American Leadership in Artificial Intelligence (in-degree 23), BIS Framework for AI Diffusion — Interim Final Rule (RESCINDED) (22), Tennessee SB 1580 / HB 1470 — AI Mental-Health Professional Impersonation Prohibition (21).
• 56 of 170 slice-7 pages carry last_updated: 2026-06-06 — the same import signature the 2026-09-13, -14, -15 and -16 reports identified in slices 3 through 6. Fifth consecutive slice. It is one import, not dozens of independent gaps, and it will keep generating candidates in every remaining slice until it is treated as one job. This is now the single largest recurring source of deferred candidates in the rotation.

Carried forward

• Citron & Chesney, Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (score 5) — second deferral. Canonical hosts already located on 2026-09-16 (SSRN abstract 3213954; Boston University repository scholarship.law.bu.edu/faculty_scholarship/640), so this is a capture away from closing. Liar's Dividend names it as a dangling origin text in its own relationships section.
• litigation/protect-democracy-v-oncd (score 5) — second deferral. No. 1:26-cv-03064. The CourtListener connector is working, so the docket is retrievable; deferred on the cap only.
• Young–Cantwell measure — closed today. Recorded here so the backlog stays honest: this was a carried-forward item and it is no longer one.
• DeepSeek-V4.1-Flash technical report (score 5) — fourth deferral. Still reported as "possible match under another address." The possible-match state needs a human look, not another scan.
• Anthropic August 2026 Risk Report — recurring false positive, second statement. The primary-doc-scan.py possible-match lane flagged it again against the 2026-09-14 08:09 digest. The vault holds it under its canonical landing page while the digest cites the CDN PDF path; this is the exact case schema v4.15 documented when the scanner was written. Not a gap, and not worth a third statement — a scanner allowlist entry would close it permanently.
• H.R. 10315 (Gottheimer, AI labs in K-12), score 6 — sixth 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) — seventh deferral; the eleven other California child-safety bills signed September 10 (score 5 as a set) — seventh deferral, and still a placement decision rather than research.
• Anthropic Institute economic scenarios working paper (score 5) — fourth deferral, a named hole in The Anthropic Institute's own open questions.
• entities/center-for-american-progress (score 4) and entities/bilal-chughtai (score 3) — second deferral.
• entities/rum-group (score 4) — fourth deferral; Chen Yixin / China Cyberspace (4) — third; South Korea's expanded espionage law (5) — fourth; Epoch AI, The Nvidia-sized hole in US GDP statistics (5) — fifth; 404 Media / Marshall Project DHS targeting teams (5) — fourth; Terence Tao, "A Severe Misalignment of AI in Mathematics" (5) — sixth; Nvidia–Groq DOJ antitrust investigation (5) — seventh; entities/bureau-of-land-management (4) — fourth.
• Ninth and tenth deferrals: entities/tech-transparency-project and the TTP CSAM-advertising report (6); the expanded Anthropic Max-tier class action (6, now unblocked by CourtListener); 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 (third deferral).
• entities/adam-thierer (score 3) — twentieth consecutive deferral, structural. Reported once, not re-listed.
• Three slice-3 pages, nineteenth 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 (seventeenth): American Civil Liberties Union (ACLU), Access Now. Slice-5 (sixteenth): Helen Toner, International Monetary Fund (IMF), Henry Farrell. Slice-6 (fifteenth): Stanford Center for Research on Foundation Models (CRFM), Sen. Elizabeth Warren, plus Simon Willison's missing ## Relationships. Slice-8: GEMA v. Suno (Munich Regional Court), Sony Music v. Udio (second suit, 2026) — both unblocked by CourtListener and both now two runs past that. Slice-9: Grok Imagine (twelfth; 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-fourth consecutive report, structural. Out of this skill's lane.
• Date-stamped headers in lint-report.md — thirty-first cycle, structural. Now 25, unchanged. Self-references 6, unchanged.

Findings that are not gaps

• Two of yesterday's top deferred candidates closed themselves overnight. litigation/umg-v-distrokid and entities/who-decides were both named in the 2026-09-16 deferred backlog — the first as "the first new-page candidate" if nothing else created it. The nightly developments-log fold created both. UMG Recordings v. DistroKid carries confidence: high on 6 sources with full litigation frontmatter. Neither was re-raised; the dedupe rule worked as written.
• The primary-document pre-filter lane produced one hit and it was already closed. primary-doc-scan.py flagged the UMG v. DistroKid complaint in the keyword lane against the 2026-09-15 22:05 digest. The complaint was verified and queued on 2026-09-16 through CourtListener. The pre-filter behaved correctly; the item was simply already in flight, which is exactly what step 3 drops.
• One digest is unprocessed and that did not stop this run. 2026-09-16-2211-ai-developments.md has no Operation: Developments-Log entry. Earlier reports deferred candidates on that ground; the 2026-09-16 report did not, and built a page from an unprocessed digest. The three pages created today all rest on unprocessed or partially-folded digests, and each is written from a primary document or the subject's own site rather than from the digest, which is what makes that safe. The nightly fold owns the prose updates to existing pages; it does not own primary-document capture.
• The [[sources/…]] placeholder in a provenance note is a broken wikilink. Prior provenance notes wrote the upgrade instruction as ` … , which lint-scan.py counts as a distinct broken target. Both pages written today were checked and the phrasing changed to a sources/ wikilink `. Worth one line because the pattern is copied forward from report to report.

Verification notes

• No broken outbound links introduced. All four created or expanded pages were resolved programmatically against the full content-page set after writing: zero unresolved wikilinks on each (the two sources/… placeholders were caught by that check and fixed).
• 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, with no exceptions to record.
• Revision fidelity. One page was edited rather than created. On California AB 3030 (Healthcare AI Disclosure), no prose sentence, figure, name, table row, citation or existing relationship line was removed; the edit renames one header and adds two lines. All five original related: entries survive verbatim.
• Confidence not raised anywhere. Three pages created at medium, never high. The expanded page's confidence and sources_count were deliberately not touched, because the edit added no source.
• sources_count set by count, not by feel. Ratepayer 10, Future of AI Innovation 2, DeepMind Institute 4 — each the number of distinct cited URLs on the page, checked after writing. The Future of AI Innovation page rests on two congress.gov URLs and says so; it is not dressed up.
• Capture completeness stated, not assumed. H.R. 9340: complete, with the EH-version caveat recorded. SCSP/Gallup: complete, all 12 pages. OpenAI framework: complete for the post, six sub-reports explicitly not captured. S. 3952: not captured, queued with the verified URL at ~79,000 characters.
• Cap respected. Eight gaps actioned against a 6–10 cap; 25 candidates after dedupe. Eleven web pulls; 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

Built Ratepayer Protection Act (H.R. 9340) from the bill text rather than the coverage — which showed the statute works through PURPA consider-and-determine rather than adoption, and that its covered class is defined without the words "artificial intelligence" or "data center"; closed the Young–Cantwell identifier gap on the second try with S. 3952, which turns out to have four sponsors and to predate the September negotiation by seven months; created DeepMind Institute; and captured OpenAI's misalignment-disclosure framework and the Gallup/SCSP sentiment report complete — with the run's standing finding being that 56 of 170 legislation/ pages carry the same 2026-06-06 import date, the fifth consecutive slice to show it, and one authenticity failure, the 19-group Senate coalition letter, which exists only inside a Washington Post exclusive.