AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-02

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

Contents

Scanned

Recent window: 4 dev-log files (2026-08-31 08:15 and 22:10; 2026-09-01 08:11 and 22:10), 79 content pages with last_updated in the last 48 hours. Rotation slice: 6 — entities/ O–Z (126 pages). Lint baseline at scan start: 212 distinct broken wikilinks, 0 referenced 2+ times; 26 date-stamped headers (down from 28); zero banned headers, self-references, standing predictions or header hyperbole; 1 unprocessed digest (2026-09-01 22:10 — the nightly fold had not yet run at scan time). primary-doc-scan.py --days 3: 6 digests, 9 cited primary-document URLs, 3 with no raw file (two already in flight as open INGEST- tasks from 2026-09-01), 2 possible matches under another address (both confirmed held), 3 lane-2 keyword hits.

Gaps actioned (7 of 24 found)

New pages created (live)

None. The two obvious new-page candidates in the live thread — a models/ page for Claude Fable 5.1 / Mythos 5.1 and litigation/isbell-v-suno — are items of the unprocessed 2026-09-01 22:10 digest, which the nightly developments-log fold owns; creating them here would collide with that run. The model page is additionally routed through the queued system-card ingest below, whose task body records the decision to be made (new pages versus Snapshot rows on Claude Fable 5 and Claude Mythos 5).

Pages expanded (live)

  • RAND Corporation — score 4 (in-degree ≥10 +2, compute/export-controls core +2); in-degree 28, the highest in slice 6 by a wide margin, on 602 words with sources_count: 2 and confidence: high — a rating the two-source count did not support. 602 → 1,244 words. Added the Technology and Security Policy Center as the AI research unit; Matheny's dated start (July 2022), IARPA, NSC/OSTP and NSCAI roles from his RAND profile; Heim's TASP compute-research lead and Pardee professorship; the Securing AI Model Weights report (RR-A2849-1, 2024; 38 attack vectors, five security levels) and its uptake as the SL2–SL5 scale in AI 2027; the AGI rise-and-fall scenarios report and the Vermeer historical-analogues report, both ingested but unlinked from the entity page; Heim's diffusion-framework Perspective and hardware-enabled-governance working paper; the Logic Theorist origin at RAND; the wargaming finding cited in Mutually Automated Destruction: The Escalating Global A.I. Arms Race (NYT). Two typed supports: links and one related: added. [3 new (Source: URL) cites; sources_count 2 → 9; confidence held at high, now supported by the count.]
  • Tom Brown — score 6 (live thread +3: named in the 2026-09-01 22:10 digest as scheduled to address the G20 ministerial; frontier-lab governance core +2; in-degree 6–9 +1). A targeted refresh: the page lacked sources_count entirely (frontmatter defect, now 10), and the graph held four facts about Brown that the page did not — co-authorship of Scaling Laws for Neural Language Models, In-context Learning and Induction Heads (Olsson et al., Anthropic, 2022) and Constitutional AI: Harmlessness from AI Feedback; the July 2026 Blomfield hire into his compute team; the August 28 Cursor compute statement. Added a G20 section from Reuters and Axios stating he was slated to speak on September 2, with the explicit note that his remarks were not yet available. 474 → 825 words. [4 new (Source: URL) cites.]

Graph wiring

None. The referenced-2+ lane was empty before and after the run; the long tail was not re-audited, per the standing finding.

Queued — foundational sources

  • Anthropic, System Card: Claude Fable 5.1 & Claude Mythos 5.1 (2026-09-01) — score 8 (live +3, uncaptured primary document +3, frontier-lab core +2). Gap type 9: "cited by source 1, not retrieved" in the 22:10 digest, and the instrument behind the day's principal release. Verified: anthropic.com landing URL → www-cdn.anthropic.com PDF (HTTP 200; 206 section headings; dateline September 1, 2026; executive-summary claims match the independently scraped launch post). Saved: Raw Sources/System Card - Claude Fable 5.1 & Claude Mythos 5.1 (2026-09-01).md (full text, 404,839 chars). Queued: INGEST-claude-fable-5-1-mythos-5-1-system-card.md. The launch post Introducing Claude Fable 5.1 and Claude Mythos 5.1 was saved alongside as a supporting raw (source_class: supporting, press-release) for the same ingest to fold under Ingest-Supporting — it carries the benchmark table, the cache-read price cut, Enterprise Frontier Safeguards, the anti-distillation change and the EU AI Act watermark.
  • OpenAI, Path to Astra: critical capabilities and frontier safeguards (2026-09-01) — score 8 (live +3, primary document +3, safety core +2). The digest had it only as a search excerpt. Verified: openai.com, HTTP 200; Wired and Axios corroborate the Critical designation, two chained zero-days, ExploitBench 100% and the August 28 RL restart; distinctive passages present verbatim. Saved: Raw Sources/Path to Astra - Critical Capabilities and Frontier Safeguards (OpenAI, 2026-09-01).md. Queued: INGEST-openai-path-to-astra-2026-09-01.md. Figures the digest lacked and the ingest should carry: 91.5% vs 59% cyber-jailbreak refusal; 56% honeypot-target access rate for GPT-5.6 Sol without safeguards versus none for Astra.
  • CNAS, Superpowers and AGI (Jacob Stokes, 2026-08-27) — score 8 (live +3 via the 2026-09-01 08:11 digest, primary document +3, national-security core +2). The digest retrieved only the executive summary. Verified: the landing page's full link list exposed the PDF on CNAS's own files.cnas.org S3 bucket (32 pages, HTTP 200); exec summary matches the PDF verbatim; cnas.org/reports and the author's own posts date it August 27. Saved: Raw Sources/Superpowers and AGI - How U.S.-China Competition Could Be Shaped by the Emergence of AGI (CNAS, Stokes, 2026-08-27).md. Queued: INGEST-cnas-superpowers-and-agi-stokes-2026.md.
  • Anthropic Alignment Science, Training a Misaligned Reward Seeker (Qi, Wright, MacDiarmid, Hubinger, August 2026) — score 8 (live +3, primary document +3, alignment core +2). Named as the "companion research post" in the 08:11 digest and linked from the ingested Anthropic, Improving our alignment and security efforts (August 2026), but never captured. Verified: alignment.anthropic.com, HTTP 200, full section structure §1–6 and appendices A.1–A.7. Saved: Raw Sources/Training a Misaligned Reward Seeker (Anthropic Alignment Science, Qi et al., 2026-08).md (203,107 chars). Queued: INGEST-anthropic-reward-seeker-2026.md.
  • SB Energy, Inc. Form S-1 (SEC EDGAR, 2026-09-01) — score 8 (live +3, primary document +3, compute core +2). Verified by structured extraction from the EDGAR document (HTTP 200): registrant, "Subject to Completion. Dated August 31, 2026", ticker SBE on Nasdaq and Nasdaq Texas, SoftBank as controlling shareholder, the "substantially dependent on OpenAI" risk-factor sentence, and the six-month figures — net loss $(3,208.9) million, revenue $138.7 million — which match the digest's rounded values. URL-only; no raw saved (large-document pattern). Queued: INGEST-sb-energy-form-s-1-2026.md, which warns that the $5–7 billion raise size is WSJ's figure, not the filing's.

Five verification records written to Wiki/_meta/queue/gap-scan/proposed-sources/.

Authenticity-verification failures

None. One dead URL: RAND's Technology and Security Policy Center page returned HTTP 404; the center's existence and Heim's role in it are sourced instead from the ingested RAND — How AGI Could Affect the Rise and Fall of Nations and the PEA3776-1 landing page already cited on Lennart Heim.

Examined and not queued

  • OpenAI, "Our agreement with the Department of War" — listed as "cited by source 1, not retrieved" in the 2026-08-31 08:15 digest. Fetched and verified (openai.com; dated February 28, 2026 with a March 2 update). Not queued: it is a February document mentioned in passing by an item about the DoD's ChatGPT Mil / GenAI.mil release, not the instrument the item is about, and no content page names it — so it scores neither the type-9 +3 nor the type-1 +3. Recorded in the deferred backlog at score 2 with the verified URL so a future run can queue it without re-fetching if a page comes to depend on it.

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

  • Live-thread items owned by the nightly fold (score 3 each). The 2026-09-01 22:10 digest is unprocessed at scan time (unprocessed_devlog=1). Its page-level work — a Fable 5.1 / Mythos 5.1 model page or Snapshot rows, litigation/isbell-v-suno (identity-rights claim against Suno, distinct from UMG Recordings v. Suno (AI music training data)), the Enterprise Frontier Safeguards retention change on Anthropic, the Astra Critical designation, the H.R. 10152 markup on Open-Source AI Leadership Act (H.R. 10152, 119th Congress) — belongs to developments-log. If the digest is still unprocessed on 2026-09-03 this becomes the second consecutive report of the same backlog.
  • "Carolina Principles" (score 3) — Kratsios did not provide a copy; Reuters has no text. Not a page until a document exists. Michael Kratsios is the current anchor.
  • OpenAI, "pacing model development" post (openai.com, score 5: named inside the verified Path to Astra text as the source for the training-pause decision, safety core +2) — new this run; not in Raw Sources/ by URL grep. One pull next run.
  • The Information's "recurrent depth" report on Astra (score 3) — paywalled abstract only; hold as attributed prose on Astra when folded, not a source to capture.
  • OpenAI "Our agreement with the Department of War" (score 2) — verified URL openai.com; see above.
  • House E&C CMT Subcommittee markup memo (score 2) — the third lane-1 hit from primary-doc-scan.py (HMTG_119_IF_17_20260901_SD_002). A committee staff memo is supporting, not foundational; the bill text it accompanies (H.R. 10152) is already captured. Left for the fold to cite as (Source: URL).
  • Slice-6 thin anchors under the corrected diagnostic (in-degree ≥6 with sources_count ≤ 2, no ## Relationships, or under 350 words): 24 of 126 pages. Worked: rand, tom-brown. Next by score: Stanford Center for Research on Foundation Models (CRFM) (in-degree 14, 485 words, sources_count: 2, high — same over-rating pattern as RAND had, score 4); Special Competitive Studies Project (SCSP) (10, 1,151 words, sources_count: 2, score 4); Simon Willison (10, no ## Relationships, score 2); Robert Chesney (in-degree 9 on 137 words, the thinnest page in the slice, score 3); Sen. Elizabeth Warren (9, 319 words, high on 2 sources, score 3); Sanjog Misra (9, 270 words, sources_count: 1, score 1); Paolo G. Carozza (9, high on 2 sources); Rishi Bommasani (8, high on 1 source); Project Liberty Institute (8, 291 words); Stanford Cyber Policy Center (7, 228 words, high on 1 source); Global Dialogue on AI Governance (7, 200 words). Five of these carry confidence: high with sources_count ≤ 2, which the decay table does not permit; a future run or a lint pass should lower them or source them.
  • Slice-5 carryover (second deferral): American Civil Liberties Union (ACLU) (24, sources_count: 2) and Access Now (20, sources_count: 2).
  • Three slice-3 pages, fourth deferral (scores 3, 2, 2): Standards-Industrial Complex, Three Theories of Victory (US / EU / China AI Governance), Transaction Costs (Coase) and the Agentic Economy.
  • entities/adam-thierer (score 3) — fifth consecutive deferral, structural. Unchanged; two mentions.
  • Twenty-one source-robustness-check candidates across slices 11–13 — tenth consecutive report, structural. Led by California SB 53 — Transparency in Frontier AI Act (in-degree 254) and TAKE IT DOWN Act — Source Summary (97).
  • models/deepseek-v3 — twelfth run, structural. Routed to model-robustness-check on 2026-08-22, unaddressed.
  • Date-stamped headers in lint-report.md (score 1) — seventeenth cycle, structural, now 26 (from 28).
  • Carried unchanged. Marvell (no company page, since 2026-08-21); California AB 1856 (text not located); the Texas OAG / Meta Webster–Mahoney exchange; the draft executive order on a frontier-developer self-regulatory organization (no public text); the AI Evaluator Forum / AEF-1 standard (score 4, raised 2026-09-01); ExploitBench flag-level disclosure (score 3).

Verification notes

  • No broken outbound links introduced. Every wikilink on the two expanded pages and the five new INGEST- tasks was resolved programmatically against the content folders; three initially wrong targets in the task bodies (sources/anthropic-risk-report-august-2026, concepts/artificial-general-intelligence, concepts/frontier-safety-frameworks) were corrected to Anthropic Risk Report: August 2026 (Redacted), Artificial General Intelligence (AGI) and AI Safety Cases and Frameworks before the run closed. One link on RAND Corporation first pointed at AI Biosecurity for a claim that lives on CBRN Uplift; corrected.
  • Style scan clean. Both expanded pages checked against the banned-header, hyperbole and self-reference lists: zero hits.
  • Revision fidelity. Every sentence, table row, relationship and "Sources in Wiki" entry of the pre-revision RAND Corporation survives; additions are either cited (Source: URL) or drawn from facts already on ingested wiki pages, each linked. Tom Brown was purely additive. One typed link was corrected during review: an initial depends-on: [[ai-2027]] on the RAND page had the dependency backwards and was rewritten as related:.
  • Confidence set honestly. RAND Corporation stays high, now backed by nine sources where the page previously claimed high on two. Tom Brown stays medium; the G20 material is scheduling, not substance.
  • Raw-source fidelity. The four raws saved are full extracted texts; figure content is described by caption in square brackets and nothing was summarised or trimmed. The announcement's 22 customer quotes are preserved in full.
  • Cap respected. Seven gaps actioned against a 6–10 cap; 24 candidates found. Eleven web pulls (two returned 400K- and 200K-character documents handled through the sandbox), two searches, one 404, no crawls.
  • Dashboard rebuilt — 2,102 article pages written to Wiki/_meta/site.nosync/. Broken-link count unchanged at 212; the referenced-2+ lane is still empty.

One-line summary

Five primary documents behind the last two days' digests — the Fable 5.1 / Mythos 5.1 system card, OpenAI's Astra Critical designation, the CNAS AGI report, Anthropic's reward-seeker paper and the SB Energy S-1 — are verified and queued for review; RAND and Tom Brown expanded live; the nightly fold owns the 2026-09-01 22:10 digest's new pages.

Categories: Briefings | briefing | gap-scan