Gap Scan — 2026-09-23

Scanned

Recent window: 3 New Developments Log/ files inside 48 hours (2026-09-21 22:03, 2026-09-22 08:07, 2026-09-22 22:10; the first was also read by yesterday's scan), 274 content pages carrying a last_updated of 2026-09-21 or 2026-09-22 — 119 of them in sources/ from yesterday's 69-paper benchmark-source ingest, 54 in evaluations/. Rotation slice: 13 — sources/ Q–Z (189 pages) + Raw Sources/ orphan check.

Run-order caveat. bin/lint-scan.py reports all three digests in the window as unprocessed — no Operation: Developments-Log entry covers them. The skill assumes this scan runs after the nightly fold; today it did not. The recent-thread lane therefore could not rely on the nightly run having created pages, and the one model page created live below (Claude Opus 5.5) is one the nightly fold would ordinarily have made. When the fold runs, it should fold into that page, not create a second one.

Lane results before prioritisation:

• Broken links. 207 distinct broken targets, 0 referenced 2+ times, 207 single-reference long tail. Seventh consecutive empty candidate lane; not re-audited.
• Primary documents. bin/primary-doc-scan.py --days 3: 6 digests, 1 cited primary URL missing (the DocumentCloud Fabrizio memo — the same false positive recorded yesterday, held under its canonical address), 7 keyword pre-filter hits. As on 09-22, the real type-9 gaps came from reading the items: the Opus 5.5 system card, OpenAI's third-party-assessment principles and the TC260 framework were all behind digest items that cited either the primary URL or only a secondary, and none had a raw file.
• Anchor thinness (slice 13). Re-scored on body-to-raw ratio at in-degree ≥ 6, as on 09-21 and 09-22. Lowest ratios: singularity-is-near-kurzweil (0.3%, in-degree 7), sb-energy-form-s-1-2026 (0.4%, 8), superintelligence-bostrom (0.4%, 14), second-machine-age-book (0.9%, 11), stanford-hai-ai-index-2026 (1,294 words against 136,054, 1.0%, in-degree 125), scientific-progress-ai-horvitz-mitchell (282 words, 1.8%, 11), remote-labor-index (412 words, 3.0%, in-degree 49). The three books are long by nature and a low ratio there is expected.
• Raw Sources/ orphan check. 1,100 markdown raws; 22 are named nowhere in Wiki/. None is an uncovered foundational document: 13 are supporting-class news captures whose wiki citations use straight rather than curly apostrophes, and the rest are companions or alternate captures of documents that already have sources/ pages (the Gemini 3.7 and 3.8 Flash "evaluation methodology" companions; DeepSeek-V3 Technical Report - full text; Inkling Model Card (Thinking Machines Lab, 2026-07-15).md alongside the cited Inkling Model Card.md — a probable raw duplicate for lint/ingest step 1.25; NIST AI Risk Management Framework (page capture 2026-07-25).md).
• Recent thread. Named actors, models, instruments and cases from the three digests checked against the graph. Genuine misses: Claude Opus 5.5 (no model page), Mercor (named on 13 content pages, no page), Eddie Wu (carried over), TC260's framework (zero mentions of any version), OpenAI's third-party-assessment post, the Trump UNGA address, Maven Smart System (11 pages, no page), Cameron Stanley (8 pages), Andy Burnham (9 pages), Stream Data Centers, Patrick Wardle. GPT-6 Sol and Luna are a fold into Astra for the nightly run, not a page gap.

Gaps actioned (7 of 38 scored; the unchanged long-tail carryover from 09-22 not re-scored)

New pages created (live)

• Claude Opus 5.5 — score 5 (live +3, core area +2). Lead item of the 2026-09-22 22:10 digest; no model page, and the nightly fold has not run. Built from Anthropic's launch post, with Reuters, TechCrunch, The Verge, VentureBeat and a developer read of the model docs: the launch benchmark table (nine benchmarks, five models, with each setting caveat — safeguard fallbacks during scoring, Zapier's no-fallback AutomationBench run, OpenAI-reported comparator figures), cost-at-effort claims, the behavioral-audit and containment results, the evaluation-awareness caveat, the three fallback safeguards, pricing and availability. confidence: medium. A > **Provenance note:** records that capability and safety claims should move to the system card once ingested, and that the benchmark rows belong on evaluations/ pages under v4.17 at that point. Depends on queued source: Claude Opus 5.5 System Card. 1,950 words, 7 sources.
• Mercor — score 4 (live +3, 13 content-page mentions with no link target +1). The 404 Media item of 2026-09-22 on contractors fired for using AI named Mercor; the company already carried the whole APEX benchmark family on APEX-Agents and GDPval and appears in two Grok model cards and the collective cyber-defense letter. Page covers business model, APEX benchmarks, the 404 Media contractor-conduct report with Mercor's statement, and a Snapshot of the October 2025 $10B Series C (company's own post), the July 2026 reported $20B talks and the August 2026 Nvidia report, and the company-stated $1B → $2B annualized revenue. Whether the $20B round closed is stated as unconfirmed. confidence: medium, 7 sources. One inbound link added on APEX-Agents.
• Eddie Wu — score 5 (live +3, carried on 09-22's deferred list, named on 7 pages +2). Built from Alibaba's own leadership page, the 2023 appointment coverage, KrASIA's account of the 2025 Apsara keynote (the three-stage path to ASI; the $4 trillion investment forecast; the RMB 380 billion plan), SCMP on the March 2026 model task force, and SCMP and Alibaba's own post on the 2026-09-22 Apsara keynote (20 GW by 2032; Zhenwu V900; Qwen roadmap). The SCMP and the digest's source disagree on whether the 5–10 trillion-parameter range applies to Qwen 5 alone or to Qwen 4.5 and 5; both are stated with attribution. confidence: medium, 7 sources. One inbound link added on Alibaba / Qwen Team.

Pages expanded (live)

• Remote Labor Index: Measuring AI Automation of Remote Work (Mazeika et al., 2025) — score 3 (in-degree 49 +2, slice 13 +1). 412 words against a 13,751-word raw at in-degree 49, the highest-in-degree thin page in the slice that could be rebuilt in one pass from its raw. Rewritten against the raw to add the full lead-author and senior-author lists and arXiv ID; project composition with the six example projects; the category distribution (video 31%, CAD 12%, graphic design 11%, …); human time and cost statistics (mean 28.9 h, median 11.5; mean $632.60, median $200; $143,991 total); the two-stage collection pipeline (64 → 43 categories; 207 + 7 + 33 projects; 358 freelancers; 550 → 240; $41 average payment); the private/public split and contamination controls; all four metrics including autoflation; the manual evaluation protocol with agreement statistics (94.4%; 56.9% ternary vs 33.0% chance); the full six-agent results table and scaffolds; the four failure-mode shares and the success categories; the authors' generalization argument and stated limitations; a ## Provenance section. The Epoch AI February 2026 comparison (What do 'economic value' benchmarks tell us? (Brand and Burnham, Epoch AI, 2026)) added as the one new cite. 412 → ~1,700 body words. Every original fact, quotation, citation and typed relationship preserved; the two interpretive sentences in the original ("claims about displacement rest on capability benchmarks never designed…"; "its value is as a series rather than a point") were re-anchored to the paper's own wording. Original backed up to Wiki/_meta/_revision-backups/remote-labor-index.2026-09-23.md. Confidence held at high, sources_count held at 1 (a primary paper).

Link aliases fixed

Zero. Seventh consecutive scan; settled finding, not re-audited.

Queued — foundational sources

• Anthropic, Claude Opus 5.5 System Card (2026-09-22) — score 8. Gap type 9, behind the lead item of the 22:10 digest. Verified: anthropic.com → www-cdn.anthropic.com (Anthropic's own CDN), PDF titled "Claude Opus 5.5 System Card", 230 pages; linked from the launch post; corroborated by Reuters, TechCrunch and The Verge. Queued URL-only: INGEST-claude-opus-5-5-system-card-2026.md; verification record in queue/gap-scan/proposed-sources/. A directed-quote extraction from the PDF failed at the scraper, so the CB-1/CB-2 determination reported in the scraper's machine summary is not on the model page and is flagged in the task for confirmation at ingest.
• OpenAI, "Priorities and principles for effective third party assessments" (2026-09-22) — score 8. Gap type 9; four priority areas and seven principles for independent assessment of frontier-AI safety cases, safeguards, Preparedness and alignment evaluations, and misalignment incidents. Verified: openai.com, byline Lama Ahmad; corroborated by PYMNTS and by openai.com's own cross-listing; distinctive passage confirmed. Saved: Raw Sources/OpenAI - Priorities and principles for effective third party assessments (2026-09-22).md (full text). Queued: INGEST-openai-third-party-assessments-principles-2026.md, with a note to ingest it alongside yesterday's queued "Building standards" post, which it links.
• TC260, AI Safety Governance Framework 3.0 (September 2026) — score 6. Gap type 9; carried via DataGuidance with no primary URL. Verified: tc260.org.cn, the committee's own site; 136-page bilingual PDF; cover, issuer and date (2026年9月) and the full table of contents read from the first three pages; corroborated by DataGuidance, AI Safety China and a public post linking the same PDF. Queued URL-only: INGEST-tc260-ai-safety-governance-framework-3-0-2026.md. No version of this framework — 1.0 (2024), 2.0 (2025) or 3.0 — is named on any content page, although TC260 is named on five; the task notes that 1.0 and 2.0 should be raised as separate gaps.

Authenticity-verification failures

None. Four documents were sought; three verified and queued.

The fourth, President Trump's 2026-09-22 address to the 81st UN General Assembly (score 8), is deferred, not failed. The White House had not posted a transcript at scan time; the UN General Debate page carries video, audio and a generated transcript, and Roll Call Factbase carries a secondary transcript. Neither is the statement as issued, so nothing was saved. Record: queue/gap-scan/proposed-sources/trump-unga-81-address-2026.md, with the hosts to re-check next run.

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

• Trump UNGA-81 address (8) — new, awaiting primary host. See above.
• stanford-hai-ai-index-2026 (score 4) — new. In-degree 125, 1,294 words against a 136,054-word raw (1.0%). The single most-relied-upon thin page in slice 13, deferred only because a faithful rewrite against a 136k-word report is a multi-pass job; it is a better fit for source-robustness-check than for this lane's one-pass expansion.
• Maven Smart System (score 4) — new. Named on 11 content pages (the Minab strike, Chief Digital and AI Office (CDAO), Palantir Technologies, Drew Cukor, Autonomous Weapons), live in the 22:10 digest, no page. Placement is unclear under "Where does X go?" — a Palantir product deployed by DOD is neither a company, a model nor an agency — so this needs a curator call before a page, and should not be auto-created.
• entities/cameron-stanley (3) and entities/andy-burnham (3) — new. 8 and 9 content-page mentions respectively; both live in the 22:10 digest.
• TC260 AI Safety Governance Frameworks 1.0 (2024) and 2.0 (2025) (4 each) — new. Raised by the 3.0 capture above.
• scientific-progress-ai-horvitz-mitchell (3), sb-energy-form-s-1-2026 (3), superintelligence-bostrom (3), second-machine-age-book (2) — slice-13 expansion candidates, new. The two books are low-ratio by nature; the Horvitz–Mitchell page at 282 words is the clearest one-pass candidate.
• Minor recent-thread items for the nightly fold, not page gaps: GPT-6 Sol and Luna (fold into Astra); Z.ai's ZCode upload incident (fold into Zhipu AI); Stream Data Centers and Anthropic's reported 1 GW lease talks; Patrick Wardle's Muse zero-day; Hochul's kill-switch remarks (AI kill switch); Bessent as AI-czar frontrunner (Scott Bessent); CFTC's extended review of CME GPU-rental futures; Nscale S-1 figures (Nscale Limited — Form S-1 registration statement (2026)).
• California data-centre bill package (5) — second deferral. The 08:07 digest adds that the signing covered air-quality, water-use and ratepayer-cost measures but still carries no bill numbers; unchanged blocker, plus the open curator placement question on California bill packages.
• Grok 4.7 and Grok 4.6 evaluation rows (4) — second deferral. Unchanged; now joined by the Opus 5.5 launch-post rows, which should move together with the card ingest.
• Carried from 09-22, each one deferral further: sources/california-ccpa-regulations (4, fourth — now structural; the worst ratio in the wiki at 1.0% against 41,547 words, in-degree 28); sources/gpt-54-thinking-system-card (4, third — structural); sources/americas-ai-action-plan (in-degree 300) and sources/california-sb-53 (277) (4 each, third — structural; raw files still unmeasured); sources/machine-platform-crowd-book, metr-frontier-risk-report-feb-mar-2026, nscale-s1-registration-statement-2026, metr-long-tasks (3–4, second); entities/data-center-coalition (3, second); Joe Benton's 2026-09-12 X thread and Substack post (5, second); entities/tech-oversight-project, companies/corridor, entities/mike-lawler (2–3, third — structural); H.R. 10315 (6, twelfth); California SB 1119 / AB 1864 and the September 10 child-safety set (6 / 5, thirteenth); OpenAI 2026–2030 financial presentation (6, unqueueable); DeepSeek-V4.1-Flash technical report (5, tenth); the three remaining OpenAI misalignment reports (4); and the unchanged carryover list in the 09-22 report (Zerra DC, CAP, Bilal Chughtai, RUM Group, Chen Yixin, South Korea espionage law, Epoch's GDP-statistics piece, TTP and its CSAM-ads report, the Max-tier class action, Bright Data, Jev, Damo Radar, Astra for Law, Muse agent app, Tau2Bench).
• Structural, reported once and not re-listed: entities/adam-thierer (26th); slice-3 Standards-Industrial Complex, Three Theories of Victory (US / EU / China AI Governance), Transformative AI (TAI) (25th); slice-4 American Civil Liberties Union (ACLU), Access Now (23rd); slice-5 Helen Toner, International Monetary Fund (IMF) (22nd); the source-robustness-check candidates across slices 11–13 (30th — remote-labor-index closed today inside this lane, the third such closure in three runs). Date-stamped headers in lint-report.md remain 14; self-references 6 (ninth cycle).

The items newly marked structural above (four slice-11/12 thin sources and three low-score entity gaps) have now missed the top 10 on three consecutive runs. They are not transient: the four sources/ pages need source-robustness-check runs, which this lane's one-pass budget does not cover, and the three entity pages will not reach the cap on score alone unless a digest item makes them live.

One-line summary

Seven gaps actioned: three pages created live (Claude Opus 5.5, Mercor, Eddie Wu), one thin anchor rebuilt from its raw (Remote Labor Index: Measuring AI Automation of Remote Work (Mazeika et al., 2025), in-degree 49), and three foundational primary documents verified and queued (the Opus 5.5 system card, OpenAI's third-party-assessment principles, China's TC260 AI Safety Governance Framework 3.0); the Trump UNGA address is waiting on a primary host, and the nightly developments-log fold has not processed the last three digests — the curator should check that job.