AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-08-30

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

Contents

Scanned

Recent window: 4 New Developments Log/ files (2026-08-28 08:10 and 22:10; 2026-08-29 08:09 and 22:04) and 59 wiki pages with last_updated or mtime inside 48 hours. Rotation slice: 3 — concepts/ P–Z (317 files in concepts/, 128 in the P–Z range).

Candidates generated: 19. Actioned: 6. Deferred: 13.

Gaps actioned (6 of 19 found)

New pages created (live)

Graph wiring

Six additive wikilink edits on existing pages, no prose rewritten:

Pages expanded (live)

None. See "The slice-3 thinness lane is mostly a false positive" below — the mechanical flags in this slice did not survive reading.

Zero, and the referenced-2+ lane is empty for the sixth consecutive run. bin/lint-scan.py reports 214 distinct broken targets across 214 references — 0 referenced 2+ times, 214 referenced once. Per step 3 the single-reference long tail is excluded from candidate generation and is not re-audited. Reported without re-auditing, as the skill directs.

Queued — foundational sources

  • H.R. 10044, AI Tax and Work Protection Act (introduced 2026-08-06) — score 8 (live thread +3, uncaptured primary document +3, US regulation core +2). The one lane-1 hit from bin/primary-doc-scan.py --days 3 that was not already dropped. Verified: congress.gov, HTTP 200, one version, sponsor Rep. Casar with Foushee and Jacobs, referred to Education and Workforce and to Ways and Means. Saved: Raw Sources/H.R. 10044 - AI Tax and Work Protection Act (119th Congress, 2026-08-06).md. Queued: INGEST-hr-10044-ai-tax-and-work-protection-act.md.
  • Pew Research Center, From Diagnoses to Treatments, Why Americans Use AI Chatbots for Health (2026-08-25) — score 6 (live thread +3, uncaptured primary document +3). Verified: pewresearch.org, HTTP 200, authors Kikuchi, Pasquini and Yam, 3,488 U.S. adults surveyed June 22–28, 2026. Saved: Raw Sources/Pew Research Center - From Diagnoses to Treatments, Why Americans Use AI Chatbots for Health (2026-08-25).md. Queued: INGEST-pew-ai-chatbots-health-2026.md.
  • Ai2 / Providence-Swedish, Surprisal-based large language models reveal immunologic insights in lobular breast cancer (2026-08-27) — score 6 (live thread +3, dangling foundational reference +3). Verified: allenai.org, HTTP 200, 27-page PDF retrieved in full; corroborated by Ai2's own announcement blog linking this exact URL. Saved: Raw Sources/Ai2 and Providence-Swedish - Surprisal-based LLMs Reveal Immunologic Insights in Lobular Breast Cancer (2026-08-27).md. Queued: INGEST-ai2-autodiscovery-lobular-breast-cancer.md.
  • GLM-5.3 License (Z.AI, August 2026) — score 6 (live thread +3, uncaptured primary document +3). Verified: served from the zai-org/GLM-5.3 repository itself, HTTP 200, bilingual text with matching operative terms. Saved: Raw Sources/GLM-5.3 License (Z.AI, 2026-08).md. Queued: INGEST-glm-5-3-license.md.

Source-fidelity findings carried into the queue tasks

Three of the four captures diverge from the digest account they came from. All three are recorded in the INGEST- task rather than fixed silently, because folding a primary source is a curated step.

  • The Ai2 paper reverses the digest's central claim. The 2026-08-29 evening digest reports that "Ai2's AutoDiscovery system found a stronger immune signature." The paper's abstract states: "The system did not identify clinically meaningful novel findings without human input." The cold-start run produced 100 hypotheses, 27 positive, and "none was judged to be clinically useful or actionable"; the finding came from a warm-start run seeded with roughly four hours of an oncologist's annotations, which the authors list among their own limitations. This is the modality/attribution class the source-fidelity skill exists to catch.
  • The Ai2 paper is not a medRxiv preprint. The digest's ## Primary documents block types it that way. It carries no medRxiv identifier, DOI or preprint-server banner; a title search returns only Ai2's blog and trade coverage. Its printed title also reads "...in lobular breast cancer," where Ai2's own blog, the coverage and the digest all drop "lobular."
  • The GLM-5.3 revenue test reaches affiliates, and the obligation is broader than hosting. The digest's transcription of the trigger is accurate. The license adds that the test runs on the Licensee "and its affiliates" on both limbs, and that a triggered licensee must pass review "before using the Software or its derivative works for any commercial purpose," not merely before hosting it.
  • H.R. 10044 is the one case where the existing page's characterisation holds up against the text. The page carried the sponsor's framing correctly; what it lacked was every operative mechanic — the 2%/3% base rates, the U-4 indexing on a three-year high-water mark, the 10^25 FLOP definitional threshold, the R&D carve-outs, and 14 of the 17 enumerated grant categories.

The slice-3 thinness lane is mostly a false positive

The mechanical anchor-thinness test (under ~700 words, or confidence: low / sources_count: 1, with in-degree ≥6) surfaced 19 pages in concepts/ P–Z. Reading the top four shows the test is misfiring on this slice, and the reason is structural rather than incidental.

  • Post-Deployment AI System Monitoring (in-degree 29, 697 words) is confidence: high on three sources, carries a six-category table, five challenge categories, and eight typed relationships. It is short because it is dense, not because it is underbuilt.
  • Specification Hazard (in-degree 16, 616 words, sources_count: 2) and Three-Lane Standards-Based AI Governance (in-degree 12, 493 words, sources_count: 3) are single-framework pages, each faithfully scoped to a named source and attributed throughout — Imas, Lee & Misra (2025) and Clearwater (March 2026) respectively. Expanding them from supporting web sources would pad them past what their anchor source says.
  • Rogue Internal Deployment (in-degree 21, 679 words) carries sources_count: 8.

concepts/ P–Z is where the vault's named-framework pages cluster, and a named-framework page is correctly bounded by its framework. Word count is the wrong instrument here; the diagnostic that would work is in-degree high, sources_count = 1, and no introduced-by or supports relationship to a sources/ page. On that test the slice yields three candidates rather than nineteen: Standards-Industrial Complex (397 words, sources_count: 1, in-degree 6), Three Theories of Victory (US / EU / China AI Governance) (1,034 words, sources_count: 1, in-degree 8) and Transaction Costs (Coase) and the Agentic Economy (735 words, sources_count: 1, in-degree 8). Deferred below rather than actioned, since none clears the cap against this run's gap-type-9 hits.

No page was expanded this run. That is a deliberate call, not a miss.

Authenticity-verification failures

None. Four documents sought, four retrieved from their canonical host on the first attempt, all four corroborated independently. No rejections, no secondary-copy fallbacks, no firecrawl failures — the first run since 2026-08-25 with a clean pull record.

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

  • entities/adam-thierer (score 3) — second consecutive deferral. Named on Federal Trade Commission (FTC) and Radical Optionality: Governing Transformative AI Under Uncertainty and the named voice in the 2026-08-28 digest's Paxton data-center item.
  • legislation/eu-cyber-resilience-act (score 3) — second consecutive deferral. The CRA's September 11, 2026 reporting layer over NIS2, GDPR and MDR vigilance; named on three content pages.
  • Three slice-3 pages under the corrected thinness test (scores 3, 2, 2) — Standards-Industrial Complex, Three Theories of Victory (US / EU / China AI Governance), Transaction Costs (Coase) and the Agentic Economy. All three are sources_count: 1 with in-degree 6–8. New this run.
  • Live-thread items owned by the nightly fold, not by this skill (score 3 each). The 2026-08-29 22:04 digest is unprocessed (unprocessed_devlog=1, confirmed by bin/devlog-backlog.py: 260 folded, 1 actionable). Its four items — the SpaceX Bastrop blades-and-vanes foundry, Amazon's agreement to acquire DuckLabs, the Ai2 AutoDiscovery result, and the Pew health survey — belong to developments-log. Two of the four had primary documents this scan captured; the items themselves are left alone.
  • companies/ducklabs (score 2) — examined and declined, not deferred. "DuckLabs" and "DuckDB" return zero mentions across all content folders. A company the vault has never referenced is not a gap in the graph; it is an item for the nightly fold to introduce, and creating a page ahead of that fold would invert the order. Revisit only if the fold names it.
  • Twenty-one source-robustness-check candidates across slices 11–13 — seventh consecutive report, structural. Unchanged, led by California SB 53 — Transparency in Frontier AI Act (in-degree 254) and TAKE IT DOWN Act — Source Summary. This is not a gap-identifier job and will not be actioned here; it needs a source-robustness-check run.
  • The 28 date-stamped headers in lint-report.md (score 1) — fourteenth cycle, structural. Unchanged in count and substance. The check does not distinguish a litigation procedural history, where chronology is the subject and the schema permits it, from a news-feed accretion. Fixing the finding means fixing the check.
  • Carried unchanged. models/deepseek-v3 (in-degree 49 on 608 words, routed to model-robustness-check on 2026-08-22, still unaddressed — ninth run, structural); the draft executive order creating a self-regulatory organization for frontier developers, which remains unactionable because no public text exists.
  • Dropped, not deferred. The Louisiana Attorney General's certified petitions against Snap and Discord, two lane-1 hits from primary-doc-scan.py. Wiki/_meta/queue/needs-review/INGEST-louisiana-ag-snap-discord-petitions-2026.md records this as a decision awaiting the curator, and step 3 excludes needs-review items from re-raising.

Run integrity

  • No broken outbound links introduced. Every wikilink on the two new pages and the six edited pages was resolved against the full page set after writing: zero unresolved.
  • Style scan clean, first pass. Both new pages checked programmatically against the banned-header, hyperbole, self-reference, standing-Predictions and date-led-header patterns. Zero hits on each.
  • Revision fidelity. All six edits to existing pages were single-word wikilink wraps around text that was already there. No sentence was rewritten, no fact or citation removed.
  • Confidence set honestly. Both new pages are confidence: medium. Neither met the three-independent-source bar: Greg Casar rests on congress.gov plus the sponsor's own press release plus positions already sourced elsewhere in the vault, and Pew Research Center on Pew's own about page plus two ingested Pew sources.
  • A queue-file wikilink was checked rather than assumed. INGEST-hr-10044 references [[concepts/compute-thresholds]]; the page exists.
  • Cap respected. Six gaps actioned against a 6–10 cap; 19 candidates found. Six scrapes and two searches, no crawls, no failures.

One-line summary

Four primary documents captured and queued for curated ingest — the H.R. 10044 bill text, the Pew health-chatbot report, the Ai2 lobular breast cancer paper and the GLM-5.3 license — with three source-fidelity divergences from the digests recorded in the queue tasks, of which the Ai2 one reverses the digest's claim about autonomous discovery; two entity pages created live; and the slice-3 thinness lane diagnosed as a structural false positive that needs a different test rather than more expansions.

Categories: Briefings | briefing | gap-scan