Gap Scan — 2026-09-22

Scanned

Recent window: 4 New Developments Log/ files (2026-09-20 08:05 and 22:08; 2026-09-21 08:30 and 22:03), 213 wiki pages carrying a last_updated of 2026-09-20 or 2026-09-21 — 51 of them in evaluations/, the continuing v4.16/v4.17 build-out. Rotation slice: 12 — sources/ I–P (203 pages).

Lane results before prioritisation:

• Broken links. bin/lint-scan.py reports 207 distinct broken targets, 0 referenced 2+ times, 207 in the single-reference long tail. The candidate lane is empty by construction for the sixth consecutive scan. Not re-audited, per the settled finding in the skill's autonomy rules.
• Primary documents. bin/primary-doc-scan.py --days 3: 6 digests, 4 cited primary-document URLs of which 1 had no raw file, 0 possible matches under another address, 9 keyword hits in the pre-filter lane. The one precise-lane hit is a false positive and is discussed below.
• Anchor thinness (slice 12). The confidence: medium + sources_count: 1 criterion again over-fires on sources/ — 37 of 203 pages at in-degree ≥ 6 — so the slice was re-scored on body length against raw-file length, as on 09-21. Worst ratios at in-degree ≥ 10: on-the-biology-of-a-llm (503 words against a 189,467-word raw, 0.3%, in-degree 45), machine-platform-crowd-book (607 against 117,183, 0.5%, in-degree 12), metr-frontier-risk-report-feb-mar-2026 (725 against 92,795, 0.8%, in-degree 11), nscale-s1-registration-statement-2026 (1,660 against 202,031, 0.8%, in-degree 12), metr-long-tasks (379 against 19,424, 2.0%, in-degree 30). kimi-k2-technical-report (14,235 words, in-degree 206) and openai-o1-system-card (22,281 words) are the slice's well-built anchors and were not flagged.
• Recent thread. Named actors, laws, models and cases from the four digests were checked against the graph. The nightly developments-log runs had already created or held pages for the Tumbler Ridge litigation, the JAWBONE Act, Irregular, Amazon v. Perplexity, the UN Scientific Panel, Alibaba/Qwen, CXMT and Meta's Muse family. Genuine misses: Joe Benton, the OpenAI standards paper, the UN panel's thematic brief, the EU data-centre rating scheme, the Data Center Coalition, Eddie Wu, and the seven California data-centre bills signed 2026-09-21.

Gaps actioned (7 of 24 found)

New pages created (live)

• Joe Benton — score 5. Named on four content pages (AI Existential Risk, Jacob Coxon, Natural Emergent Misalignment from Reward Hacking in Production RL (MacDiarmid, Wright, Uesato et al., Anthropic), Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026)) and carried as a live item in the 2026-09-20 08:05 digest, with no page. Built from his September 12 X thread as quoted by three outlets, the Reuters interview of September 19, the MATS program listing of his Anthropic role, and his co-authorship of six papers already held here (Sabotage Evaluations for Frontier Models, on which he is first author; Many-shot Jailbreaking; Constitutional Classifiers; Chain-of-Thought Monitorability; Natural Emergent Misalignment from Reward Hacking; Automated Weak-to-Strong Researcher). confidence: medium — his title and team are reported inconsistently, and his own primary texts are not captured. 993 words, 9 sources.

Pages expanded (live)

• models/grok — score 7. The page's lead, infobox and lineage table all stopped at Grok 4.6 while SpaceXAI released Grok 4.7 on 2026-09-21. Added a ### Grok 4.7 capabilities subsection with the developer's full eight-row comparison table (CursorBench 4.0, DeepSWE v1.1, EEBench, AA Briefcase v1.1, Terminal-Bench 4.0, Harvey Legal Agent Benchmark, HealthBench Professional, plus token prices) and its GDPval Elo figures; a training paragraph (new larger base model, longer RL run weighted toward multi-hour tasks, native Grok Bot harness); a safety paragraph recording the first specific safety figures SpaceXAI has published for a Grok release (LatchBio biosafety 62.4%, HackerBench v0.3 at 3.3% of risky dual-use prompts allowed through) together with the fact that there is still no system card and no named external evaluator; and availability and pricing. Lead, infobox and lineage table updated. 2,217 → 2,885 words, 1 new (Source: URL) cite, sources_count 10 → 11.
• On the Biology of a Large Language Model — score 4. The worst body-to-raw ratio in the slice and in the wiki: 503 words against a 189,467-word raw, at in-degree 45. The original listed nine case studies in a sentence each and stopped. Rewritten against the raw to add the full author list from the paper's own BibTeX (28 authors, Lindsey through Batson); the methodology section (cross-layer transcoders, perturbation testing, polysemanticity as the motivating problem, the authors' stated ~25% success rate and their "existence proofs" framing); concrete detail for every case study the original named, including the "Michael Batkin"/pickleball hallucination example, the "Babies Outlive Mustard Block" jailbreak, and the reward-model-bias model organism developed by Sam Marks and Johannes Treutlein; a new ## Structural observations section carrying the paper's ten named graph-structure findings; a new ## Limitations section carrying the six case-study failure modes and seven methodological issues the authors enumerate; a new ## Conclusions the authors draw section carrying the seven properties named in § 15.1, including the authors' statement that complexity is their most consistent finding; and a ## Provenance section. 503 → 2,440 words (ratio 0.3% → 1.3%). Every original fact, citation and typed relationship preserved; original backed up to Wiki/_meta/_revision-backups/on-the-biology-of-a-llm.2026-09-22.md. Confidence held at medium — still a single source.
• AI Existential Risk — correction, not an expansion. The page stated that Benton's resignation "and his reasons for it became public on September 19, 2026 when he told Reuters." Research for the Benton page shows he left in late August and published his reasons himself on X on September 12; the Reuters interview came a week later. Sentence rewritten with the September 12 statement and its quotations, the Reuters interview kept and correctly dated, the METR move added. sources_count 9 → 10.

Link aliases fixed

Zero. Sixth consecutive scan with no fixable alias: the 207 broken targets are all single-reference author surnames, institution shorthand and deliberate forward references. Four bare links on sources/on-the-biology-of-a-llm were normalised to full folder paths during the rewrite ([[emergent-misalignment]] → [[concepts/emergent-misalignment]] and three others), but the lint report confirms bare slugs already resolved, so these were cosmetic and are not counted as alias fixes.

Queued — foundational sources

• OpenAI, "Building standards for the next phase of AI" (2026-09-21) — score 8. Gap type 9; leads the 09-21 22:03 digest and named on zero content pages. An original position paper arguing the United States should lead international technical standards for frontier AI including recursive self-improvement, through CAISI and the network of AI safety institutes. Verified: openai.com, the company's own domain; corroborated by Axios; distinctive passage confirmed ("Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely"). Saved: Raw Sources/OpenAI - Building standards for the next phase of AI (2026-09-21).md (full text). Queued: INGEST-openai-building-standards-next-phase-ai-2026.md.
• Independent International Scientific Panel on AI, AI Agents, Misalignment and the Risk of Losing Human Control: Evidence from the OpenAI-Hugging Face Incident (2026-09-21) — score 8. Gap type 9; carried in the 09-21 22:03 digest through The Verge. The Panel's first thematic brief and the first intergovernmental-level assessment of an incident the vault already covers from the company disclosures and the METR investigation. Verified: un.org, the Panel's own subsite; PDF fetched and read; cover confirms "ADVANCE UNEDITED VERSION 1 • 21 SEPTEMBER 2026" and the Panel's personal-capacity disclaimer; corroborated by UN News story 1168380 and by the press release and media advisory on the same host. Queued URL-only: INGEST-un-scientific-panel-thematic-brief-ai-agents-2026.md; verification record in queue/gap-scan/proposed-sources/.
• Commission Delegated Regulation establishing a common Union rating scheme for data centres, C(2026) 3472, and annexes (2026-09-21) — score 6. Gap type 9; carried in the 09-21 22:03 digest through Reuters with no primary URL. The first EU-level data-centre instrument to reach the vault, against five US and Canadian ones already in legislation/. Verified: energy.ec.europa.eu, DG Energy's own publication page and PDF paths; corroborated by Commission press release IP/26/1667 and the DG Energy topic page; empowerment (Article 33(3) of the recast EED) and the amended Regulation (EU) 2024/1364 both resolve on EUR-Lex. Queued URL-only: INGEST-eu-data-centre-rating-scheme-delegated-regulation-2026.md; verification record in queue/gap-scan/proposed-sources/, including a flag that a placeholder-bearing pre-adoption draft of this text circulates on a non-EU host and must not be used as the raw source.

Authenticity-verification failures

None. Three documents were sought and all three verified on their canonical hosts.

One near-miss is worth recording. A third-party PDF of an earlier, undated draft of the EU delegated regulation is indexed on a Korean standards-notification site and still carries the Commission's own unfilled placeholders — "data centres in Europe account for around \[x%\] of its total electricity consumption, \[y%\] of its total water consumption." It is a genuine draft rather than a fabrication, but it is not the instrument, and a fetch that resolved to it instead of the europa.eu PDFs would have put bracketed placeholders into the graph as figures. The rejection is noted in the verification record so the ingest does not repeat the search.

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

• The seven California data-centre bills signed 2026-09-21 (score 5) — new this run. Newsom signed a package requiring operators to disclose electricity use, water consumption, land use and workforce needs and to comply with state energy-procurement requirements. The digest names no bill numbers, so this needs a research step before it can be queued, and it lands on the same California-bill-package placement question the curator already holds open from SB 1119, AB 1864 and the September 10 child-safety set.
• Grok 4.7 evaluation rows for evaluations/ (score 4) — new this run, deliberately not actioned. Under v4.17 the Terminal-Bench 4.0, HealthBench Professional and GDPval figures added to Grok (xAI) today belong on their evaluations/ pages rather than inline. They were left inline to match the existing Grok 4.6 treatment on the same page and, more importantly, to avoid writing into evaluations/ while the system-card-pass operation is live on the other machine — the collision risk flagged in the 2026-09-20 report. A single follow-on pass should move both the 4.6 and 4.7 tables together; splitting them now would leave the page inconsistent with itself.
• entities/eddie-wu (score 4) — new this run. Alibaba's chief executive is named on seven content pages and has no page, while Alibaba / Qwen Team carries his announcements. His 2026-09-22 Apsara Conference remarks (a 5–10 trillion parameter model, Qwen 4 in training, the Zhenwu V900 chip, a 20 GW capacity target for 2032) are Snapshot work for the nightly fold, not a page gap; the missing person page is.
• entities/data-center-coalition (score 3) — new this run. Quoted in the 09-21 22:03 digest opposing the California package, with zero mentions across all content pages. A trade association recurring in data-centre siting politics, but below the cap on score alone.
• Joe Benton's X thread and Substack post of 2026-09-12 (score 5) — new this run. The new Joe Benton page rests entirely on secondary accounts of these two primary texts. A foundational capture would replace three Indian-outlet citations with the author's own words. Recorded in the page's ## Open questions as well.
• sources/machine-platform-crowd-book (0.5% ratio), metr-frontier-risk-report-feb-mar-2026 (0.8%), nscale-s1-registration-statement-2026 (0.8%), metr-long-tasks (2.0%) — slice-12 expansion candidates, score 3–4 each, new this run. All below on-the-biology-of-a-llm on the combined ratio-and-in-degree measure; deferred on the cap, not on merit. The METR long-tasks page at in-degree 30 is the strongest of the four.
• Slice-11 carryover, second and third deferrals. sources/california-ccpa-regulations (score 4, third — still the worst ratio in the wiki at 1.0% against a 41,547-word raw, in-degree 28); sources/gpt-54-thinking-system-card (4, second); sources/americas-ai-action-plan (in-degree 300) and sources/california-sb-53 (in-degree 277), 4 each, second — their raw files still have not been measured, and whether they are underbuilt or merely short relative to their sources needs that check before either is actioned.
• entities/tech-oversight-project (2), companies/corridor (3), entities/mike-lawler (2) — second deferral each. All three remain genuine type-3 gaps below the cap on score.
• Arizona HB 2311 bill text — dropped, in flight. INGEST-arizona-hb-2311-bill-text-2026.md is open in the queue. Yesterday's deferral is now correctly excluded as work in flight rather than deferred again.
• H.R. 10315 (Gottheimer, AI labs in K-12), score 6 — eleventh deferral. Standing caveat unchanged: the digest cites an all-actions URL, not bill text.
• California SB 1119 and AB 1864 bill texts (6 each) and the eleven other California child-safety bills signed September 10 (5 as a set) — twelfth deferral. The placement question is a curator decision, not research. Today's data-centre package joins the same question.
• OpenAI's 2026–2030 financial presentation (6) — structurally unqueueable, fourth report carrying it. Seen by the Financial Times and not public. Snapshot-table work for OpenAI, not a source pull. The 09-21 08:30 digest's "Reported but unverified" item on a $278 billion cash-burn forecast, disputed the same day by a person identifying himself as an OpenAI product finance director, is the same thread and is equally unqueueable.
• DeepSeek-V4.1-Flash technical report (5) — ninth deferral. Still "possible match under another address."
• The three remaining OpenAI misalignment reports at alignment.openai.com (4). Unchanged; three of six are queued individually.
• Carryover, unchanged: entities/zerra-dc (3); entities/center-for-american-progress (4) and entities/bilal-chughtai (3); entities/rum-group (4); Chen Yixin / China Cyberspace (4); South Korea's expanded espionage law (5); Epoch AI, The Nvidia-sized hole in US GDP statistics (5); entities/tech-transparency-project and the TTP CSAM-advertising report (6); the expanded Anthropic Max-tier class action (6); Bright Data (3); models/jev (3); Damo Radar (3); OpenAI Astra for Law (3); Meta's Muse agent app (3); a Tau2Bench page (4). Anthropic's Life Sciences Verification Program (5), Figure Helix 2.5 and Crusoe Series F (4 each) remain raw-capture gaps rather than page gaps, all three subjects having pages.
• Structural, reported once and not re-listed: entities/adam-thierer (25th); the three slice-3 concept pages Standards-Industrial Complex, Three Theories of Victory (US / EU / China AI Governance) and Transformative AI (TAI) (24th); slice-4 American Civil Liberties Union (ACLU) and Access Now (22nd); slice-5 Helen Toner and International Monetary Fund (IMF) (21st); the 21 source-robustness-check candidates across slices 11–13 (29th — though two of them, claude-sonnet-46-system-card on 09-21 and on-the-biology-of-a-llm today, have now been closed inside this skill's expand lane, which suggests the "out of lane" framing has been overstated).
• Date-stamped headers in lint-report.md — thirty-fourth cycle, structural. Now 14, down from 17. Self-references 6, unchanged for an eighth cycle.

The primary-document scanner's URL-matching defect now extends past Anthropic

The precise lane produced one hit and it is a false positive: the DocumentCloud URL for the Fabrizio memo cited in the 09-20 08:05 digest, documentcloud.org/documents/28624377-messaging-memo-neutralizing-data-center-attacks/. The memo was pulled and verified by yesterday's scan and is held as Raw Sources/Fabrizio Lee - Neutralizing Political Attacks on AI Data-Center Development Through Enforceable Conditions (2026-09-14).md, under its canonical Fabrizio Lee address rather than the DocumentCloud mirror the news story linked.

This is the same defect the 09-19 and 09-20 reports recorded for the Anthropic Advanced AI Framework and the August 2026 Risk Report — the scanner matches on URL string, so a document held under one address is invisible when a digest cites another — but it is no longer an Anthropic-CDN problem. It is the general case: news coverage routinely links a mirror (DocumentCloud, Scribd, a court-document host) while the vault holds the primary. A normalisation table keyed only on www-cdn.anthropic.com and cdn.sanity.io would not have caught today's hit. What would is a title-similarity check against Raw Sources/ filenames before a URL is declared missing. That is a bin/ change and ships alone.

Separately, the scanner's own baseline is now three consecutive runs in which the precise lane produced more false positives than real gaps. Today's three actioned document captures all came from the keyword pre-filter lane and the recent-thread scan, not from the precise lane — the OpenAI standards paper and the UN brief because the digest cited the primary URL alongside the coverage and the scanner therefore saw nothing missing about them, and the EU regulation because the digest cited only Reuters and the instrument was found by reading the item. That is the lane working as designed, and it is worth saying plainly: the pre-filter lane, which the skill describes as needing judgment, is currently carrying this gap type.

One-line summary

Seven gaps actioned: one page created (Joe Benton), two expanded and one corrected, three foundational primary documents verified and queued (OpenAI's standards paper, the UN panel's first thematic brief, the EU data-centre rating regulation); the run's substantive finding is that the Benton item as folded from the digest carried the wrong date and the wrong channel for his going public, and that the primary-document scanner's URL-matching defect has now produced a false positive outside the Anthropic CDN case it was first diagnosed on.