AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-15

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 48h (2026-09-13 22:05, 2026-09-14 08:09, 2026-09-14 22:05 — the last of these unprocessed at scan time) and 67 wiki pages with a modification time inside the window. Rotation slice: 5 — entities/ H–N, 148 pages.

Scans run: bin/lint-scan.py (authoritative broken-link report), bin/primary-doc-scan.py --days 3, dangling-reference scan over the recent window, anchor-thinness scan over the slice, the open queue and the last seven gap-scan reports for dedupe.

Gaps actioned (6 of 23 found)

New pages created (live)

One divergence recorded on the page rather than resolved. The vault's Jack Clark page says Clark "heads the Anthropic Institute," sourced to a Cosmos Institute announcement of 2026-04-30. Contemporaneous reporting on the June 2026 essay consistently identifies Marina Favaro as the Institute's head or lead. The two are not necessarily inconsistent — Clark is Anthropic's policy head — but no source reconciles them, and Anthropic's own institute landing page is client-rendered and returned no readable content on attempted capture. Both attributions are carried with their sources, under a > **Provenance note:**, and the question is listed under ## Open questions.

Pages expanded (live)

The rotation slice returned 25 thin anchors (under 350 words at in-degree ≥6) out of 148 pages — the second productive thin-anchor slice in a row, after three consecutive slices returning zero or near-zero through 2026-09-13. One was expanded.

  • Jay Obernolte — score 5 (in-degree 9, +1; US AI regulation core, +2; +2 for the specific gap below). 153 words, sources_count: 2, last_updated: 2026-06-06, for the member of Congress who has authored or co-authored the two largest bipartisan House AI legislative products of the 119th Congress. The page recorded only the April 2026 Lieu–Obernolte bill.

What was missing was the larger half of his record. Added the Great American AI Act discussion draft of 2026-06-04 with Rep. Lori Trahan (D-MA) — 269 pages, its preemption of state laws "specifically regulating the development" of AI models on a three-year sunset expiring December 2029, its preservation of state laws of general applicability and of state authority over post-deployment activity, its workforce-data provisions, its incorporation of H.R. 3460 (the AI Whistleblower Protection Act, which Obernolte sponsored), and the divided House reaction including Trahan's conditional assent ("I will preempt, but only if we're setting the strongest possible federal standard… A preemption provision in a bill with no meaningful federal standard is a moratorium") and Rep. Josh Gottheimer's judgment that the draft "cannot serve as the basis for productive dialogue." Added his January 2026 signalling of the bill as a vehicle to codify the Center for AI Standards and Innovation, his House Energy and Commerce membership, and his Republican Policy Committee chairmanship. 153 → 971 words; 12 new (Source: URL) cites; sources_count 2 → 6; three new typed relationships. confidence held at medium.

One divergence recorded rather than smoothed. The page said Obernolte "co-chaired" the House Bipartisan AI Task Force. The Task Force's own final report lists him as "Jay Obernolte (CA-23), Chair"; DLA Piper describes him as having "served as a co-chair"; Tech Policy Press refers to another member as "Obernolte's co-chair." All three are recorded with their sources. The original "co-chaired" claim is preserved, not overwritten.

Zero, without re-auditing the list. bin/lint-scan.py reports 0 broken targets referenced 2+ times; all 212 distinct broken targets are single-reference long tail, which step 3 places permanently out of scope. This is the seventh consecutive scan to report zero on the settled 2026-08-18 basis, and the finding needs no further confirmation.

One near-miss resolved and recorded so it stops presenting: legislation/great-american-ai-act is not missing. The Act is covered at Frontier Act / Great American AI Act (Obernolte–Trahan), which the alias-resolution step located before any page was proposed. This was the run's most plausible-looking missing-legislation candidate and it resolved to an existing page under a different name.

Queued — foundational sources

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

  • UK AI Security Institute, Our evaluation of Claude Mythos Preview's cyber capabilities (score 8, the run's highest) — gap type 1, and the item the 2026-09-14 report named "the strongest single candidate for the next run." It is the only external evidence adduced by Could rogue agent swarms take over the entire internet in the next six months? (Marcus, Hamiel and Korman, September 2026), and reaches the vault solely as a quotation inside that essay, which does not link it. Verified: aisi.gov.uk, the agency's own domain, corroborated by two later AISI posts that link back to it and independently restate the end-to-end result, by the AISI blog index, by AISI's own institutional social accounts, and by an independent trade report. Saved complete. Queued: INGEST-aisi-mythos-preview-cyber-evaluation-2026.md.

The capture changes a claim the vault currently carries. The Marcus essay renders AISI's conclusion as Mythos could "autonomously compromise [only] small, weakly defended vulnerable systems." The published sentence is: the model is "at least capable of autonomously attacking small, weakly defended and vulnerable enterprise systems where access to a network has been gained." Four differences, all running the same way: compromise for attacking; enterprise dropped; the network-access precondition dropped; and — the consequential one — "at least capable of" is a floor, and the interpolated "[only]" converts it into a ceiling. AISI's own next sentence says it "cannot say for sure whether Mythos Preview would be able to attack well-defended systems," and gives the reason: its ranges have no active defenders, no defensive tooling and no penalty for triggering alerts. That is an absence of evidence, not a finding of inability. The Marcus essay discloses the bracket as its own insertion, so nothing was concealed — but the reversal survives into any page quoting the essay rather than the source.

  • Microsoft AI, Humanist AI Code of Conduct for MAI Models (score 7) — gap type 9, and the one deferred item with a known publication date, which the last report said to check first. It published on schedule. Verified: microsoft.ai, the company's own domain, corroborated by its own announcement post (captured complete), by the Reuters report the vault already cites in the 2026-09-14 morning digest, and by two further same-day trade reports. Queued: INGEST-microsoft-mai-code-of-conduct-2026.md.

Capture is partial and the raw file says so at its head. The Code runs to roughly 103,800 characters. The raw carries the announcement post complete, including its ten-point summary panel with quotations, plus a verified structural extract of the Code — scope and status, the Chain of Command, the Absolute Constraints, the Human Control Requirements, and the Part 3 and Part 4 section structure through 4.5. The child-safety and manipulation-at-scale constraint text, the Glossary, the Objectives section and Part 4.5 onward are not captured, and the ingest task repeats the instruction to fetch the full document.

Three things about the document bear on pages the vault already holds. It is a draft that is not in force — "we are not using it to train our models today" — so describing MAI models as governed by it would state as current what the document puts in the future. Its scope is narrower than "Microsoft's AI": it binds MAI Models and says it "does not extend to other models simply because Microsoft uses or hosts them," which matters given Microsoft's hosting of OpenAI models. And two of its commitments are counter-positions to threads already tracked: the flat rejection of model welfare and AI legal personhood, against Anthropic Model Welfare (team); and a stated capability ceiling — it "rejects the race to produce an all-purpose superintelligence that could evade these safeguards… even if that means compromising on ultimate generality, autonomy, or capability" — published two days after the 2026-09-12 pacing debate. Whether the second is responsive to that debate is not stated in the document and is not asserted.

The AI content is two passages doing different work, and merging them would lose that: a governance sentence ("work faster to evolve a consensus-based global AI governance framework") in the security paragraph, and the first of five industrial initiatives — an "open source and inclusive AI initiative" under which China will "be a pioneer in establishing a BRICS AI open source community." The second carries no funding figure, institution, timetable or compute commitment, and nothing in the text says the community exists. It is also a national statement at Session II, not an agreed BRICS instrument, and binds no other member.

  • New Mexico Supreme Court, contempt order in State v. Sandoval, No. S-1-SC-40845 (2026-09-09) (score 7) — gap type 9, first-lane hit against the 2026-09-14 morning digest, which carries the item entirely through one trade account. AI-Fabricated Citations was edited inside the window, and this is the case in that line where the fabrication extended to witness testimony in a criminal proceeding rather than to case citations. Queued: INGEST-nm-supreme-court-aarons-contempt-order-2026.md.

Authenticity satisfied; capture deliberately not taken. The order is nonprecedential by its own terms, under Rule 12-405(B) NMRA, and no copy on a court-controlled host was located — both reachable copies are aggregator-hosted PDFs of the court's own order. Its distinctive text was extracted and matches six independent accounts at paragraph level (¶ 3 the ChatGPT admission, ¶ 4 the four fabricated witnesses, ¶ 12 the Disciplinary Board referral, ¶ 18 the $5,000 sanction, ¶ 19 the Rule 12-405(B) disposition), so this is not a protocol rejection. But the extraction lost the case names in both misrepresented-authority citations and the signature block, and saving a partial from a secondary host is a worse evidentiary record than an honest gap. The ingest is told to try nmonesource.com and the NM Courts case lookup first.

Two distinctions the coverage collapses and the order does not: four witnesses were wholly fabricated, while false testimony was separately attributed to three real witnesses; and Aarons did not cite fake cases — he described real ones inaccurately, a different failure mode from the Mata v. Avianca line the concept page is built around.

Authenticity-verification failures

None. Four documents sought, four verified. One capture was withheld on provenance grounds rather than rejected on authenticity grounds, and the distinction is recorded in its verification record.

Two recurring scan hits correctly excluded rather than re-raised:

  • The Liccardo letter to Speaker Mike Johnson, recorded as a curator decision in needs-review/2026-09-12-liccardo-letter-unreachable.md. Step 3 drops it. It will keep appearing until a signed version is released.
  • The Anthropic Redacted Risk Report, August 2026, flagged again by the scanner as a "possible match under another address." The vault holds it as Raw Sources/Anthropic - Risk Report August 2026 (Redacted) (2026-08-14).md; the digest cited the CDN PDF path. This is the exact false positive the v4.15 changelog anticipated, and it is working as designed.

One possible-match entry checked and confirmed as a real gap: the DeepSeek-V4.1-Flash technical report at a Hugging Face blob path. DeepSeek-V4.1-Flash now exists — created by the nightly fold since the last scan, which closes the page gap the last report deferred — but no V4.1-Flash primary document is held under any address. Deferred below.

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

New this run

  • State Bar of New Mexico ethics advisory opinion on generative AI, 2024 (score 5). First-lane scan hit, sbnm.org — the State Bar's own domain, a primary host, unlike the court order it accompanies. It is the professional-responsibility standard the contempt order operates against. Named in the ingest task above as worth capturing alongside; deferred here so it is not lost if that ingest is narrowed.
  • arXiv:2603.11214, AISI's paper describing 'The Last Ones' cyber range (score 5). Surfaced by today's AISI capture, which cites it as the detailed description of the 32-step range behind the run's headline result. Named in the ingest task; a separate foundational candidate with a resolvable arXiv identifier.
  • entities/cosmos-institute (score 3) — newly promoted out of the long tail by this run's own edit. It was a single-reference broken target, which step 3 excludes from re-audit by rule. The Anthropic Institute, created today, adds a second reference — the 2026-04-30 Cosmos Institute announcement that is the only public source for the Institute's establishment. It will therefore appear in the referenced-2+ candidate lane on the next lint-scan.py run, which has been empty for seven consecutive scans. The link was kept rather than removed: it is a real organization the vault names, and it is the provenance for a leadership claim.
  • Schiff–Banks "Collaboration on Adversarial Threats and Security Risks" and the Young–Cantwell "Future of AI Innovation Act" (score 5 as a pair). Both named in the 2026-09-14 22:05 digest with no bill number and no text URL recorded, which is why neither is actioned. The first would create an antitrust safe harbor for labs coordinating on cybersecurity and loss-of-control scenarios; the second would formally establish the Center for AI Standards and Innovation and is separately identified in a June 2026 legal commentary as S. 4178 of the 118th Congress with a 119th-Congress successor at S. 3952. legislation/future-of-ai-innovation-act does not exist and entities/jim-banks and entities/todd-young do not exist. This is the week's most consequential live legislative thread and it is blocked on identifiers, not on judgment.
  • Chen Yixin article on AI risks, China Cyberspace magazine, 2026-09-13 (score 4). The PRC State Security Minister's warning that adversaries' abuse of AI could "directly threaten China's political security, institutional security and ideological security." Named in the 2026-09-14 evening digest with no public URL recorded. entities/chen-yixin does not exist.
  • Twenty-four further slice-5 thin anchors (score 2–4 as a set). Beyond the one expanded, the slice holds 24 more pages under 350 words at in-degree ≥6, led by International Monetary Fund (IMF) (338 words, in-degree 17), Josh Shapiro (190, 11), Mariano-Florentino Cuéllar (337, 10), Hudson Institute (300, 10), Maria Cantwell (257, 10) and John Thune (296, 9). Sixteen of the 25 carry last_updated: 2026-06-06 — the same single-import date that the 2026-09-13 and 2026-09-14 reports identified behind the slice-3 and slice-4 stragglers. Three slices running have now produced the same signature, which is stronger evidence for one import than for dozens of independent gaps.
  • Maria Cantwell and John Thune specifically (score 7 and 6) — deferred on collision, not on merit. Both are thin anchors and central to the week's live thread: the September 23 Senate Commerce markup Cruz is pressing for, the Klobuchar–Thune duty-of-care draft, and Cantwell's concerns with that framework. They scored highest of any expansion candidate. They are deferred because the 2026-09-14 22:05 digest carrying that thread is unprocessed, and the nightly developments-log fold owns it first — the same ground on which the last report deferred the DeepSeek item. Expanding them today would have raced the fold on the same pages the same night. They should be the first expansion candidates next run, after the fold has landed.

Carried forward

  • H.R. 10315 (Gottheimer, AI labs in K-12 schools), score 6 — fourth deferral. First-lane hit again today. The standing caveat is why it keeps deferring: the digest cites an all-actions URL, which is not bill text, and the bill was introduced 2026-09-08. Connects to AI in Education.
  • DeepSeek-V4.1-Flash release post and technical report (score 5) — second deferral, now narrowed. The page gap closed (DeepSeek-V4.1-Flash exists); the document gap did not. Release post at deepseek.com, technical report at a Hugging Face blob path.
  • Anthropic Institute economic scenarios working paper (score 5) — second deferral, and now a named hole in a page that exists. anthropic.com. The Anthropic Institute was created today and records under ## Open questions that this paper is published but not held, with no account of its contents.
  • Epoch AI, The Nvidia-sized hole in US GDP statistics (score 5) — third deferral.
  • South Korea's expanded espionage law (score 5) — second deferral. Still no URL recorded in the coverage.
  • 404 Media / Marshall Project, DHS Predictive Intelligence Targeting Teams (score 5) — second deferral. Predictive Policing and CBP — Customs and Border Protection (AI Deployer) were both edited inside the window again.
  • California SB 1119 and AB 1864 bill texts (score 6 each) — fifth deferral. Text-capture gap; leginfo URLs are in the 2026-09-10 22:05 digest.
  • The eleven other California child-safety bills signed September 10 (score 5 as a set) — fifth deferral. The placement question — a page apiece, or one page for the package — is a decision, not research, and the cap will keep deferring it until someone makes it.
  • Terence Tao, "A Severe Misalignment of AI in Mathematics" (score 5) — fourth deferral.
  • Nvidia–Groq DOJ antitrust investigation (score 5) — fifth deferral.
  • entities/rum-group (score 4) and entities/bureau-of-land-management (score 4) — second deferral each, both held at the quality gate on sourcing, for the reasons the last report gave.
  • Seventh and eighth deferrals: entities/tech-transparency-project and the TTP CSAM-advertising report (score 6); the expanded Anthropic Max-tier class action (score 6, still blocked on an unrecorded court and case number); legislation/ pages for the AI OVERWATCH Act, Chip Security Act and MATCH Act (score 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/adam-thierer (score 3) — eighteenth consecutive deferral, structural. Reported once, not re-listed.
  • Three slice-3 pages, seventeenth 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. Today's slice-5 finding — sixteen of 25 thin anchors sharing last_updated: 2026-06-06 — is the third independent confirmation that these are not stragglers but the visible edge of one import.
  • Slice carryover, unchanged. Slice-4 (fifteenth): American Civil Liberties Union (ACLU), Access Now. Slice-5 (fourteenth, and audited directly this run): Helen Toner, International Monetary Fund (IMF), Henry Farrell — all three confirmed still below the cap, with Toner also among the slice's seven confidence: high pages on sources_count ≤ 1. Slice-6 (thirteenth): 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). Slice-9: Grok Imagine (tenth deferral; the standing recommendation that it belongs to model-robustness-check is unchanged). 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-second consecutive report, structural. Out of this skill's lane.
  • Date-stamped headers in lint-report.md — twenty-ninth cycle, structural. Now 25, unchanged from last run. Self-references holding at 6.

Findings that are not gaps

  • companies/nvidia does not exist, and that is correct. Nvidia is covered at Nvidia & TSMC — AI Compute Infrastructure — 7,622 words, sources_count: 95, in-degree 660 — which the alias-resolution step located before any page was proposed. Recorded once so the absence of a companies/nvidia slug stops reading as a gap in future runs.
  • The Great American AI Act is covered, at Frontier Act / Great American AI Act (Obernolte–Trahan). See the alias section above.
  • Slice-5 graph wiring is complete. 0 of 148 pages lack a ## Relationships section at in-degree ≥6, against 3 of 87 in slice 3 and the wiring work slice 4 required. Two consecutive healthy slices on this measure.
  • The confidence: high on thin sourcing pattern is present but small outside sources/. Slice 5 returns 7 of 148 pages at confidence: high with sources_count ≤ 1 — Herbert A. Simon, Jony Ive, Helen Toner, Katrina Manson, MIT Media Lab, Matteo Wong and Lawrence Lessig — against 5 of 205 in slice 4 and zero in slices 1–3. This is consistent with the standing finding that the pattern is heaviest in sources/ (102 of 172 and 90 of 149 in slices 12 and 13) without being confined to it. Recorded as a third data point, not as a new recommendation.
  • An in-degree counting caveat, stated once. A grep for [[entities/jay-obernolte across Wiki/ returns 31 matches; the content-page in-degree is 9. The difference is Wiki/_meta/ — log entries, briefings and reports. In-degree figures in this report are content-only, computed with _meta/ excluded, which is the correct basis for the thin-anchor rule. A future run that counts without excluding _meta/ will overstate in-degree for any page this skill has previously written about, which is exactly the set of pages it is most likely to re-examine.

Verification notes

  • No broken outbound links introduced, with one deliberate exception. Both pages were resolved programmatically against the full content-page set after writing: zero unresolved wikilinks on Jay Obernolte; one on The Anthropic Institute (entities/cosmos-institute), kept deliberately and reported above as a candidate the next lint run will surface.
  • Style scan clean. Both pages checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each. The Obernolte page's new sections are named for their subjects ("Lieu–Obernolte bill (April 2026)", "Great American AI Act discussion draft (June 2026)") and carry dates because the drafts are dated, not as changelog entries.
  • Revision fidelity. No prose sentence, figure, name, quotation, citation or existing section was removed from Jay Obernolte. Every fact and both citations in the original survive; the one claim that is now contested — "co-chaired" the AI Task Force — is preserved verbatim alongside the two sources that describe the role differently, so the change is auditable rather than silent. Both existing bare wikilinks ([[lieu-obernolte-ai-bill]], [[nist-ai-rmf]]) were left as they were, since converting bare links is a standing curator decision recorded in needs-review/2026-08-22-ambiguous-bare-wikilinks.md.
  • Confidence not raised. Obernolte held at medium despite rising from 2 to 6 sources: the new sources are a sponsor's own press release and four legal or trade commentaries on a discussion draft, which do not reach the independent-three-source bar for a high on a live legislative subject. The new Institute page was created at medium, never high, per the autonomy rule.
  • A schema question raised rather than decided. The Anthropic Institute was created with entity_type: think-tank. The two existing Anthropic-division pages, Anthropic Interpretability Team and Anthropic Model Welfare (team), both use entity_type: nonprofit, which is factually wrong for a division of a for-profit company. The enum has no value for a company research division. Rather than propagate a false value or retype existing pages on a convention question, the divergence is recorded in needs-review/2026-09-15-anthropic-institute-entity-type.md with three options for the curator. This leaves three sibling pages carrying two values, which is stated plainly because it is worse than either choice alone.
  • Source fidelity. Fifteen divergences between primary text and secondary account were recorded across the three captures, in each raw file's ## Source-fidelity notes. The three that most change a claim are the AISI modality reversal (floor read as ceiling), the Microsoft Code's not-in-force status and MAI-only scope, and the BRICS statement being a national statement rather than an agreed instrument.
  • Capture completeness stated, not assumed. AISI: complete, with the three charts represented by their published captions and the arXiv range paper deliberately not fetched. Xi BRICS: complete, with the summit photograph not reproduced. Microsoft: partial and marked as such at the head of the raw file, with the omissions enumerated. NM Supreme Court: not captured, with the reason recorded in its verification record.
  • CourtListener MCP unauthenticated for the twelfth consecutive run. It is present in the session but requires an OAuth flow this session cannot run. That directly cost this run: the NM Supreme Court order is exactly the document type the connector exists to retrieve, and the fallback was two aggregator PDFs. Two slice-8 litigation gaps and the Anthropic Max-tier class action remain blocked on it. Twelve runs is well past a transient failure and needs the curator to authorize the connector in an interactive session.
  • Cap respected. Six gaps actioned against a 6–10 cap; 23 candidates after dedupe. Four scrapes and four searches; no crawls.
  • 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

Captured the AISI Mythos evaluation the vault had been citing second-hand — and found that the quotation it carries reverses the source's modality — plus Microsoft's Code of Conduct on its publication day and Xi's BRICS statement; created The Anthropic Institute and rebuilt Jay Obernolte from 153 to 971 words; four ingests await review, and the CourtListener connector is now twelve runs unauthenticated and has started costing captures.

Categories: Briefings | briefing | gap-scan