AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-13

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

Contents

Scanned

Recent window: 4 New Developments Log/ files (2026-09-11 08:05 and 22:05; 2026-09-12 08:05 and 22:05) and 88 wiki pages touched inside the last 48 hours. Rotation slice: 3 — concepts/ P–Z (87 pages).

bin/lint-scan.py: 2,087 content files scanned; broken_distinct=212, of which 0 referenced 2+ times and 212 referenced once. banned_headers=0, predictions=0, header_hype=0, date_headers=26, self_refs=6, unprocessed_devlog=1.

bin/primary-doc-scan.py --days 3: 6 digests, 9 cited primary-document URLs, 3 with no raw file, 0 possible matches under another address, 6 second-lane keyword hits.

26 candidates survived deduplication. 6 actioned; 20 deferred.

Gaps actioned (6 of 26 found)

New pages created (live)

None. No gap this run met the AUTO-new-page conditions. The run's three strongest candidates were all primary documents — a declaration, a letter and an executive order — which belong in the QUEUE lane by rule, since folding a primary source into the graph stays a curated step. The two entity candidates the recent window produced (entities/ms-isac, entities/alex-sobel) were held at the quality gate as placement questions rather than research gaps; see Deferred.

One candidate closed rather than deferred: entities/terence-tao, carried at score 4 by the 2026-09-12 report and deliberately not created under the quality gate, now exists — created by the nightly developments-log fold since that scan. The judgment to wait for the declaration ingest rather than build the page from one blog post and one signature held up.

Pages expanded (live)

The rotation slice returned zero thin anchors (no page under 350 words at in-degree ≥6) — the second consecutive slice to do so. Three of 87 pages lacked a ## Relationships section, and those three are the three highest-in-degree pages in the slice. All three were wired this run; the slice is now complete on that measure.

  • Scaling Laws — score 4 (in-degree 136; compute core +2). Added 14 typed links. This page is also the slice's sharpest length-versus-importance mismatch: 806 words at in-degree 136, above the 350-word thin-anchor bar but the second-shortest of the slice's twenty highest-in-degree pages. Its last_updated was 2026-07-26 before this run. Recorded as an observation for the curator rather than actioned — deepening it is a source-robustness-scale job on eleven sources, not graph wiring.

No prose, figure, table row, citation or existing section was changed or removed on any of the three. Each edit is purely additive apart from one last_updated value.

Zero, without re-auditing. The lint report returns 0 broken targets referenced 2+ times — the candidate lane is empty, not merely unproductive. All 212 distinct broken targets are single-reference long tail, which the skill's step-3 exclusion places permanently out of scope: author surnames, institution shorthand, and deliberate forward references. This is the fifth consecutive scan to report zero and the finding is settled.

Queued — foundational sources

  • Leiden Declaration on Artificial Intelligence and Mathematics (score 6) — gap type 1, and the highest-value dangling reference found this run. Named by name on three content pages — Ten Advances in Mathematics and Theoretical Computer Science, Machine-produced mathematics: claims, verification and credit and Astra — plus the 2026-08-02 and 2026-09-12 digests, with no Raw Sources/ file. The wiki has carried it for six weeks entirely through OpenAI's one-sentence characterisation of it. Verified: leidendeclaration.ai, corroborated by the IMU endorsement letter on mathunion.org and Zenodo deposit 20302944. Saved: Raw Sources/Leiden Declaration on Artificial Intelligence and Mathematics (2026).md (complete: preamble, five values, five threats, all four recommendation sections, working group, endorsements). Queued: INGEST-leiden-declaration-ai-mathematics-2026.md.
  • Letter from 71 UK parliamentarians to Prime Minister Andy Burnham (score 8, the run's highest) — gap type 9, first-lane hit against the 2026-09-12 22:05 digest, which is unprocessed at scan time. The letter is the instrument the item is about. Verified: images.controlai.com, corroborated by Guardian and ControlAI accounts. Saved: Raw Sources/Letter from 71 UK parliamentarians to Prime Minister Andy Burnham on superintelligence (2026-09-11).md (complete, 3 pages, full signature block). Queued: INGEST-uk-parliamentarians-letter-burnham-superintelligence-2026.md.
  • Massachusetts Executive Order 654 (score 6) — gap type 1, carried at score 6 from the 2026-09-12 report, which named it the clear next capture in the Massachusetts thread. Referenced three times across two documents the vault already holds. Verified: the Commonwealth's own law library at mass.gov, identifier Mass Register No. 1571, corroborated by the law-library index sequence and two independent accounts of the signing date. Saved: Raw Sources/Massachusetts Executive Order 654 - To Secure Massachusetts' Energy Future (2026-03-16).md (complete: twelve WHEREAS clauses, ordering clause, Sections 1–5). Queued: INGEST-massachusetts-eo-654-energy-supply-plan.md.

Verification records for all three are in Wiki/_meta/queue/gap-scan/proposed-sources/.

Authenticity-verification failures

None. Three sources sought, three verified and saved. No document was rejected this run.

One capture deferred rather than failed: the Epoch AI report Huawei's Roadmap to 2031 (score 8) fetched successfully from epoch.ai but returned 108,812 characters, past what this session could read back and transcribe inside the remaining budget. Nothing about the document is in doubt — the host is Epoch's own publications path and the fetch returned HTTP 200. It is deferred as a capture-cost item, not a verification item, and is the strongest single candidate for the next run.

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

New this run

  • Epoch AI, Huawei's Roadmap to 2031 (score 8). Compute and export controls are a wiki core area and the report's central claim is quantitative and checkable: Huawei's most powerful chip at roughly half an H100's arithmetic performance; 0.9 million H100-equivalents produced in 2026 against Nvidia's 23 million, under 4% of Nvidia's AI compute; the conclusion that export controls constrain both levers Huawei could use to close the gap. The vault holds two Huawei raws (CSIS, SemiAnalysis) and two Epoch raws, none of them this. Deferred on capture cost only — see above.
  • Gary Marcus, Nathan Hamiel and Zack Korman, joint response to Amodei's botnet forecast (score 5). Published 2026-09-12, the same day as the essay it disputes; We Must Pace the Frontier (Amodei, September 2026) was ingested this week and this is the first substantive published rebuttal to it. It is a foundational counter-position, not news-cycle commentary: it argues the six-month agent-swarm claim is too vague to assess and implausible as stated, and cites a UK AI Security Institute finding that Anthropic's Mythos model could autonomously compromise only small, weakly defended systems. The authors explicitly do not dismiss the underlying concern. A contradicts edge from the Amodei source page has nothing to point at until this is captured. garymarcus.substack.com
  • Epoch AI, Announcing FrontierMath Erdős (score 6) and Epoch AI, The Nvidia-sized hole in US GDP statistics (score 5). Both named in the 2026-09-12 22:05 digest's ## Primary documents list. The first is the benchmark behind the claim that Astra became the first model to solve any of 68 unsolved Erdős problems formalized in Lean — a figure now on Astra with no benchmark definition behind it, which is the pattern AI Benchmarks and Evaluation exists to prevent.
  • H.R. 10315 (Gottheimer, AI labs in K-12 schools), score 6. A first-lane primary-doc-scan hit: the digest cites congress.gov with no raw file. Note that an all-actions URL is not bill text; introduced 2026-09-08, so text may not yet be posted. Connects to AI in Education, edited inside the window.
  • OpenAI, Daybreak for Frontline Defenders (score 5) and entities/ms-isac (score 3). The MS-ISAC pilot, the $1 billion subsidised-access pledge and the 2,000 approved organisations are all on the record with no page behind the programme and no entity page for the Multi-State Information Sharing and Analysis Center. The OpenAI page is a product announcement and reads as supporting, not foundational — it would fold into OpenAI and a new government or entity page rather than earning a sources/ page. Recorded so the placement question is answered once.
  • entities/alex-sobel (score 3) — a placement question, not a research gap. He introduced the UK ASI bill and is signatory 1 on the letter captured today. Whether a backbencher who has carried one bill warrants an entity page, or folds into UK Artificial Superintelligence Bill (ControlAI draft, May 28, 2026; introduced September 8, 2026) and ControlAI, is left to the ingest. The same letter's signature block also surfaces the Canadian half of the ControlAI coalition ("30+ Canadian lawmakers"), which has no page and no source behind it.

Carried forward

  • California SB 1119 and AB 1864 bill texts (score 6 each) — third deferral. Both pages exist; this is a text-capture gap. The leginfo URLs are in the 2026-09-10 22:05 digest's ## Primary documents list.
  • The eleven other California child-safety bills signed September 10 (score 5 as a set) — third deferral. The placement question (a page apiece, or one page for the September 10 package) is still undecided and is what the cap keeps deferring. It will not resolve itself; it needs a decision.
  • Stop Rogue AI Act bill text — the gap is the host, not the task. INGEST-stop-rogue-ai-act-text-2026.md is open and the cited URL is an Inside AI Policy mirror. Congress.gov should be tried first at ingest time. Third run running.
  • Fourth and fifth deferrals: entities/tech-transparency-project and the TTP report on AI-generated CSAM advertising (score 6, now fifth); the expanded Anthropic Max-tier class action (score 6, fifth, 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, fifth, placement still 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. The Terence Tao blog post (score 5, second deferral) stays open, though its associated entity gap closed this run.
  • entities/adam-thierer (score 3) — sixteenth consecutive deferral, structural. Reported once, not re-listed.
  • Three slice-3 pages, fifteenth deferral — structural: Standards-Industrial Complex, Three Theories of Victory (US / EU / China AI Governance), Transaction Costs (Coase) and the Agentic Economy. This run audited slice 3 directly and confirms the profile that keeps them below the cap: all three are confidence: medium on sources_count: 1, last touched 2026-06-06, at in-degree 6, 12 and 10. They are not thin by word count (397, 1,034 and 735 words) and so never trip the thin-anchor rule; they are under-sourced, which is source-robustness-check work, not this skill's. Fifteen runs of deferral is the system reporting that correctly. They need a direct instruction or a decision that they are not gaps.
  • Date-stamped headers in lint-report.md — twenty-seventh cycle, structural. Holding at 26. Self-references holding at 6.
  • Carried unchanged. entities/nightingale; legislation/caadca; entities/david-feith and entities/hudson-institute; Marvell; California AB 1856; the Texas OAG / Meta Webster–Mahoney exchange; the two chain-of-thought monitoring papers (arXiv:2507.11473, arXiv:2503.11926); the OpenAI June 2026 Threat Report PDF; the Epoch AI AI-R&D task taxonomy; the ARC Prize Astra evaluation post; Supreme Court of India, Vijay Ghanshyam Gadiya v. Union of India; litigation/worker-info-exchange-v-uber; entities/center-for-shared-ai-prosperity; CMORG "Firm Guidance For Frontier AI"; NCSC frontier-AI guidance; the Minnesota AG Emerging Technology report; ESRB/2026/3; the Governor of the Bank of England's 9 July 2026 letter to Dame Meg Hillier MP.

Dropped rather than deferred

  • The Liccardo letter to Speaker Mike Johnson — a first-lane primary-doc-scan hit again today, and correctly excluded: it is recorded as a curator decision in needs-review/2026-09-12-liccardo-letter-unreachable.md. That is a question waiting on the curator, not a gap to re-raise. Noted so the recurring scan hit does not read as a missed item.
  • California SB 813 — a first-lane hit whose raw file exists under a different filename (California SB 813 - enrolled text as signed (Chapter 179, 2026-09-09).md). The scanner matches on URL, and the vault captured this from leginfo rather than from the LegiScan address the digest cited. Same class of false positive the v4.15 changelog anticipated; worth a possible match under another address entry, which the scanner reported as 0 this run.

Slice-3 health, for comparison against prior slices

Slice 3 (concepts/ P–Z, 87 pages) is healthy on the structural measures and weakest on sourcing depth.

  • ## Relationships: 3 of 87 lacking, all three fixed this run — the slice is complete. Against 3 of 44 in slice 2, 18 of 191 in slice 1, 5 of 168 in slice 0. The three missing were the slice's top three by in-degree, which is the pattern worth noting: graph wiring has been weakest precisely where it matters most, presumably because the oldest and most-linked pages predate the convention.
  • Thin anchors: zero, for the second consecutive slice.
  • confidence: high on sources_count ≤ 1: zero across all 87 pages. This is the fourth consecutive slice confirming the pattern is a sources/ phenomenon rather than a wiki-wide one (102 of 172 and 90 of 149 in slices 12 and 13). The recommendation that a sources/-specific confidence rule belongs in the schema was stated for the last time as a recommendation in the 2026-09-12 report; this run is a fourth independent confirmation and nothing further is added.
  • confidence: medium on sources_count ≤ 1 at in-degree ≥6: 5 of 87 — the three structural carryovers plus Seeing Like a State (applied to AI) and Two Conditions and Three Weapons of AI Manipulation. All five carry last_updated: 2026-06-06, the same date, which suggests a single import rather than five independent gaps.

Verification notes

  • No broken outbound links introduced. All three touched pages were resolved programmatically against the full 2,087-page set after writing: 47 new wikilinks across the three ## Relationships sections, zero unresolved. Two candidate targets were checked and rejected during drafting rather than shipped — concepts/ai-2027 (the page is sources/ai-2027) and entities/jason-jia-xi-wu (does not exist; Wu is named in prose and via Techno-Federalism: How Regulatory Fragmentation Shapes the U.S.-China AI Race instead).
  • Style scan clean. All three touched pages checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each.
  • Revision fidelity. All three edits are purely additive apart from one last_updated value each. No prose sentence, figure, table row, citation or existing section was changed or removed. Existing ## See also and ## Relationship to other concepts sections were left exactly as found, including the bare unqualified wikilinks that are the subject of the standing needs-review note 2026-08-22-ambiguous-bare-wikilinks.md; the new sections carry folder-qualified targets alongside them rather than replacing them, following the precedent set on 2026-09-07 and 2026-09-12.
  • Confidence not inflated anywhere. No confidence or sources_count value was changed on any page, because graph wiring adds no evidence. No new page was created, so no new-page confidence judgment arose.
  • Source fidelity. Twenty-two divergences between primary text and secondary account were recorded across the three captures, in each raw file's ## Source-fidelity notes. The four that most change a claim: the Leiden Declaration's authorship provision is addressed to individual mathematicians, not to companies, so OpenAI's much-cited response answers a provision not written for it, and three wiki pages currently frame the declaration through that reading; the UK letter's "130+ UK parliamentarians" is the standing ControlAI coalition, a different quantity from its 71 signatories, and the two must not be merged; Massachusetts EO 654 never uses the words "data center," "artificial intelligence" or "AI" and is an energy order whose relevance is wholly derivative, so a page presenting it as data center policy would import a false premise; and EO 654 Section 5 disclaims any cause of action or right to judicial review, which adds a third layer to the already-recorded finding that the data center Framework's binding force is derivative.
  • One date recorded rather than asserted. The Leiden Declaration carries no publication date on its face; published: 2026-06 is inferred from the IMU endorsement letter's path and from TechCrunch's account, and the document itself says only that it reflects the field "as of May 2026." The raw file and the verification record both say so, and the ingest task carries finding a dated primary statement as a task.
  • One passed deadline flagged. EO 654 Section 1 required DOER to publish a review of demand and customer supply programs by September 1, 2026. Whether it did is not established by the order and is recorded as an ingest-time check rather than assumed.
  • Capture completeness stated, not assumed. Leiden Declaration: preamble, values, threats, all four recommendation sections, working group, provenance, institutional and eight featured endorsements — complete; the 3,964-name signatory register is not captured and is noted as absent, since the site renders only the latest 24. UK letter: 3 of 3 pages, full body and all 71 signatories enumerated — complete. EO 654: twelve WHEREAS clauses, ordering clause, Sections 1–5 — complete. Nothing was summarised or trimmed in any of the three.
  • CourtListener MCP unauthenticated for the tenth consecutive run. The connector is present but requires an OAuth flow this session cannot run. Two slice-8 litigation gaps and the Anthropic Max-tier class action remain blocked on it. Ten runs is past the point where this is a transient failure.
  • Cap respected. Six gaps actioned against a 6–10 cap; 26 candidates after dedupe. Four scrapes and one search; no crawls. One over-large fetch, recorded above as a deferral rather than a failure.
  • 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

Closed the run's strongest dangling reference — the Leiden Declaration, named on three content pages for six weeks and carried entirely through OpenAI's one-sentence account of it, which turns out to answer a provision the declaration addresses to mathematicians rather than to companies — alongside the 71-signatory UK letter to the Prime Minister and Massachusetts EO 654, the capture the last report named as next; and finished the rotation slice's graph wiring on the three highest-in-degree pages in concepts/ P–Z, which were also the only three in 87 lacking typed relationships. Needing the curator's attention are the Epoch AI Huawei report, deferred on capture size rather than any doubt about it and now the strongest outstanding candidate; the first published rebuttal to the Amodei essay ingested this week, which the graph has no edge for until it is captured; the still-undecided placement of the eleven California child-safety bills, now on its third deferral without a decision; and a CourtListener connector unauthenticated for the tenth consecutive run.

Categories: Briefings | briefing | gap-scan