Gap Scan — 2026-08-19

Scanned

Recent window: 4 New Developments Log/ files (2026-08-17 08:08 and 22:05; 2026-08-18 08:05 and 22:05) and 25 wiki pages with last_updated in the last 48 hours. Rotation slice: 6 — entities/ O–Z (119 pages).

Broken-link analysis ran python3 bin/lint-scan.py, which reported 244 distinct broken targets across 270 refs, of which 26 are referenced 2+ times. Only the 2+ list was worked, per the skill.

Dedupe note. All 26 of the referenced-2+ targets appear on a Deferred backlog in a gap-scan-*.md report from the last 7 days. Read literally, the step-3 dedupe rule ("listed as actioned or deferred in any report from the last 7 days") would drop every candidate and leave the broken-link lane empty, which contradicts step 4's statement that deferred items "naturally re-surface next run." The rules were reconciled in favour of step 4: actioned items were dropped as done; deferred items were treated as eligible. This is worth settling in the skill text — see the one-line summary.

Alias fixes: zero. Consistent with the settled finding recorded for the four scans through 2026-08-18. Spot-checking the 26 targets against the full slug set confirmed the pattern holds: they are author surnames (entities/sacha-altay, entities/joseph-bernstein, entities/lee-anne-fennell), institution shorthand (entities/uw, entities/ecb, entities/sais, entities/carnegie-mellon — which is not entities/carnegie-endowment), and deliberate forward references marked "(planned)" in the citing page. No target resolved to an existing page under another name.

Gaps actioned (9 of 42 found)

New pages created (live)

• Pennsylvania Executive Order 2026-05 (Data Centers) — Pennsylvania Executive Order 2026-05, signed 2026-08-18, led the 2026-08-18 22:05 brief. entities/josh-shapiro existed; the order did not. Scored +3 live thread, +2 core area (data-center siting / state AI policy). Built from the Governor's own press release, which reproduces the operative directives, plus NBC10 and NBC News for the framing quote. Includes a comparison with New York Data-Center Moratorium (2026) — the two are different instruments (conditional permitting vs. abeyance), and the page says so rather than lumping them.
• U.S.-China Economic and Security Review Commission (USCC) — U.S.-China Economic and Security Review Commission. Named substantively in the 2026-08-18 08:05 brief for its report on China's data commercialization; no page existed and no keyword match anywhere in the wiki. Scored +3 live thread, +2 core area (US-China / national security). Built from uscc.gov/about-us for mandate and composition, the Commission's own report landing page, Reuters via U.S. News, and a trade-press account. Filed under entities/ per the "Where does X go?" rule: an advisory body that studies and recommends, not an agency that deploys or a regulator that rules.
• Alondra Nelson — marked "(planned)" in sources/farahany-class-22-us-ai-policy-myths both in prose and under ## Relationships; referenced twice; architect of the Blueprint for an AI Bill of Rights. Scored +2 core area (US AI policy), +1 explicit planned reference. Built from her Institute for Advanced Study faculty page, the TIME100 AI listing, and her published AI Action Summit remarks. Confidence held at medium — the biographical spine is a first-party self-descriptive page.

Pages expanded (live)

• Seth Lazar — thin anchor: in-degree 9, 207 words, confidence: medium, sources_count: 1. Expanded to four sources. Corrected a stale affiliation: the page described him as a philosopher at ANU, but MINT Lab has relocated to the Johns Hopkins University School of Government and Policy, where he is now Professor. Added the political-philosophy-of-algorithmic-power strand (Stanford Tanner Lectures 2023; "Governing the Algorithmic City" in Philosophy & Public Affairs 2025; Kamm Lecture 2022), the normative-competence preprint linking to Normative Competence, the editorial roles, and the Knight First Amendment Institute senior AI adviser appointment (January 2024). Identified the previously anonymous "Nelson (2023)" co-author as Alondra Nelson, created the same run. 3 new (Source: URL) cites. Every prior fact and citation retained.
• Vector Institute — thin anchor: in-degree 8, 153 words, confidence: low, sources_count: 1. Expanded to four sources with a founding-and-funding section: March 2017 launch, $135M in five-year commitments from the federal and Ontario governments and ~40 industry partners, founders Frey/Hinton/Urtasun, Hinton as chief scientific adviser, and the institute's role under the Pan-Canadian AI Strategy. Confidence raised low → medium. 3 new cites. The two funding figures reported on different bases ($135M committed vs. CAD $200M received) are flagged as unreconciled rather than silently averaged.
• Susie Wiles — thin anchor: in-degree 11 (highest in the slice), 330 words. Added the direct quotes from her May 6, 2026 X post via The Hill, Hassett's same-day "just like an FDA drug" remark, Politico's placement of Wiles and Bessent in the voluntary-first-look camp, and a new "Engagement with frontier labs" section covering the Anthropic CEO meeting, Altman's late-July meeting, and the completed White House frontier AI evaluation framework (30-day voluntary pre-deployment cyber review, briefed 2026-08-04, extended to open-weight models 2026-08-12). 4 new cites. Existing "Confidence" section on the Wiles–Bessent division of labour retained unchanged.

Link aliases fixed

None. See the alias note under Scanned.

Queued — foundational sources

• OpenAI, "Pacing model development in an era of cyber-critical capabilities" (2026-08-18) — the primary company record behind the training-slowdown thread the 2026-08-18 22:05 brief leads with, and the announced successor to OpenAI Preparedness Framework V.2. Neither a sources/ page nor a Raw Sources/ file existed. Verified: openai.com (canonical company host), live fetch with maxAge: 0, full text, corroborated by Axios and TIME both dated 2026-08-18. Saved: Raw Sources/OpenAI - Pacing model development in an era of cyber-critical capabilities (openai.com, 2026-08-18).md. Queued: INGEST-openai-pacing-model-development.md. The ingest task carries a fidelity note: the article says Astra may meet the Critical cyber threshold on preliminary evidence — the modality must survive the fold.
• Anthropic, "Redacted Risk Report: August 2026" — carried under Reported but unverified in New Developments Log/2026-08-17-2205, which recorded that the PDF was retrieved but could not be read. Verified as to identity and host: landing page on anthropic.com, PDF on www-cdn.anthropic.com, table of contents and executive-summary tables confirm the document, February 2026 predecessor exists in the same series, SecureBio has published a review of an earlier edition. Content not read — it is a long redacted PDF, so per the QUEUE lane no raw was saved and the ingest fetches it fresh. Queued: INGEST-anthropic-redacted-risk-report-aug-2026.md with fetch_at_ingest: true. The widely repeated figures (≈50,000 contractors, ≈133 million exchanges, 1,197 transcripts flagged, near-one-year duration) come from downstream summaries only and are flagged in the task as unverified against the primary text.

Authenticity-verification failures

None. Both queued sources verified to their canonical first-party hosts. The Anthropic report is recorded as verified as to identity and host with content unread — a scope limit, not a failure.

Deferred to needs-review

• concepts/algorithmic-decisionmaking — a genuine missing-page gap (2 refs, both marked "(planned)"), held at the quality gate because concepts/admt and concepts/algorithmic-accountability already cover most of the ground and a third general page risks the duplication reconciled on 2026-08-15 for model-welfare. Three curator options written up in Wiki/_meta/queue/gap-scan/needs-review/2026-08-19-algorithmic-decisionmaking-vs-admt.md.

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

Recent-thread items with no page, all scoring 3–5:

• government/house-office-of-legislative-counsel — the AI-generated-bill-text repair story (2026-08-17 Politico; 61 attorneys, 5,623 drafting requests in 60 days, up 72%). Score 4.
• entities/jakob-pachocki — OpenAI chief scientist, referenced from index.md, companies/openai, concepts/risk-based-ai-regulation and 5 sources/ pages; quoted in the 2026-08-18 slowdown coverage. Score 4.
• White House frontier AI evaluation framework — a page candidate, but the administration does not plan to publish its details, so the evidence base is entirely secondary. Score 4, held on the quality gate.
• companies/worldclaw — Hong Kong venture accepting World Liberty Financial's USD1 stablecoin; 43 of 90 hosted models from flagged Chinese firms. Single source. Score 3.
• companies/noda-ai — $100M DoD award; carried as unverified, neither award date nor contract terms confirmed. Score 3.
• companies/fivetran, entities/disruptive — named in the a16z antitrust inquiry and the Groq round respectively. Score 3 each.
• ChatGPT for Teens — a product, not obviously a page; likely folds into an existing teen-safety or companion concept page. Score 3.

Broken-link 2+ candidates not reached (score 2–3 each): entities/kevin-klyman, entities/michael-j-d-vermeer, concepts/talent-flow-china-us, concepts/data-broker-regulation, entities/carnegie-mellon, entities/santa-fe-institute, entities/bank-of-england, entities/ecb, entities/federal-reserve, entities/clayton-christensen, entities/sacha-altay, entities/joseph-bernstein, entities/lee-anne-fennell, entities/adam-smith, entities/dina-powell-mccormick, entities/ilhan-scheer, entities/samuel-weinbach, concepts/economic-possibilities-for-artificial-intelligence, concepts/ai-and-language-models, concepts/a-vision-of-democratic-ai.

concepts/inverse-cooking-problem and concepts/inverse-trust-problem remain parked: reviewed 2026-07-10 as coined terms, not re-litigated here. entities/uw and entities/sais are institution shorthand with no resolvable target — note that entities/johns-hopkins, which would have absorbed entities/sais, appears in an earlier scan's actioned list but does not exist on disk; worth a look, since Seth Lazar's move makes a Johns Hopkins page more useful than it was.

Slice-6 thin anchor not actioned: entities/sanjog-misra (in-degree 7, 270 words). entities/sen-warren was dropped from the thin list — in-degree 8 but already confidence: high with 2 sources.

One-line summary

Three pages created and three expanded live, two foundational sources verified and queued for review, zero alias fixes and zero verification failures; two things need the curator: the algorithmic-decisionmaking / admt duplication call in needs-review/, and a decision on the step-3 dedupe rule, which as written would drop every deferred candidate and empty the broken-link lane.