AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-08-25

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

Contents

Scanned

Recent window: 4 dev-log files (2026-08-23 08:11 and 22:12; 2026-08-24 08:11 and 22:05), 72 pages with last_updated in the last 48 hours, of which 60 are content pages. Rotation slice: 12 — sources/ I–P (152 pages).

Mechanical scans run: bin/lint-scan.py (broken wikilinks, style residue) and bin/primary-doc-scan.py --days 3 (gap type 9). Candidates found: 34. Actioned: 7.

Wiki/_meta/lint-report.md reports 215 distinct broken targets across 215 refs — 0 referenced 2+ times. The previous ten runs reported 23–25 in that lane and deferred it each time as structural. The lane did not empty by rule change: bin/lint-scan.py is unmodified since 2026-08-18, and an independent re-derivation run this morning agrees (233 distinct / 234 refs, with the single 2+ target being _meta/briefings/weekly-2026-W19, an operational path rather than a content page). Spot-checking the twelve author-entity targets that dominated the lane confirms the cause: entities/kevin-klyman and entities/carnegie-mellon now exist as pages. The backlog was worked, not reclassified.

The consequence for this run is that the broken-link lane generated zero candidates, and the candidate list came entirely from the recent-thread scan, the dangling-reference scan, the rotation slice and the deferred backlog.

Gaps actioned (7 of 34 found)

New pages created (live)

Pages expanded (live)

None. No slice-12 page met the expand lane's terms; see the routing note below.

Zero — fifth consecutive run. The lane is now empty for a stronger reason than before: there are no referenced-2+ broken targets left to alias-resolve, and the 215-item long tail is excluded by rule.

Queued — foundational sources

  • Generalist Team, "GEN-1.5: Embodied Foundation Models are One-Shot Learners" (2026-08-19) — score 6 (live thread +3, dangling foundational reference +3). Carried as a new development in the 2026-08-23 08:11 digest, which cited the post directly; named on Generalist and two other pages, with no Raw Sources/ file. Verified: generalistai.com (developer's own domain; the page's own citation block and BibTeX corroborate title, attribution and date). Saved: Raw Sources/GEN-1.5 - Embodied Foundation Models are One-Shot Learners (Generalist AI, 2026-08-19).md. Queued: INGEST-generalist-gen-1-5-2026.md.
  • Joel Hron & Jonathan Schwarz, "Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best" (2026-07-31) — score 6 (live thread +3, dangling foundational reference +3). The primary company record for the Thomson model's training approach, content sourcing and benchmark claims; the wiki reached it only through SiliconAngle's account of the August 24 deployment. Verified: thomsonreuters.com (company domain; page metadata names both authors and the July 31 date; the company's own August 20 press release links to it). Saved: Raw Sources/Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best (Hron and Schwarz, 2026-07-31).md. Queued: INGEST-thomson-reuters-thomson-model-2026.md.

Authenticity-verification failures

None. Both queued sources came from their developer's or publisher's own live domain, in full body text rather than excerpt, with title, attribution and date corroborated independently. Verification records: Wiki/_meta/queue/gap-scan/proposed-sources/generalist-gen-1-5-2026.md and …/thomson-reuters-thomson-model-2026.md.

Two provenance limits are flagged in those records rather than suppressed:

  • Firecrawl returned connector errors on every scrape and search call after the first three. Both queued captures, and the AIA and Thomson Reuters corporate pages behind two of the new content pages, were fetched with the web_fetch fallback. That is the same failure the 2026-08-24 22:05 digest recorded. The captures are full-text and verified; the point is that their provenance is the fallback path, not a firecrawl pull.
  • Two primary sources could not be reached at all. ndia.org/about is behind an Incapsula challenge that blocked both web_fetch and a stealth-proxy firecrawl scrape; enfabrica.net/company returned an empty body. National Defense Industrial Association therefore takes its organizational facts from a search-result summary of ndia.org/about corroborated against Wikipedia, and Enfabrica takes its founding facts from trade press. Both pages are confidence: medium for that reason, and neither claim rests on a single unverified source.

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

  • Slice-12 routing: ten source-robustness-check candidates (score 4 each). The rotation slice's thinness scan surfaced sources/ pages with high in-degree and thin bodies against a full Raw Sources/ file — work that belongs to source-robustness-check, not to gap-identifier's expand lane, which is for non-source anchors. Ranked by in-degree against body length: On the Biology of a Large Language Model (in-degree 22, 3,773 body chars, medium, 1 source — the strongest candidate, and it also carries five bare wikilinks of the kind logged in needs-review/2026-08-22-ambiguous-bare-wikilinks.md); OpenAI Model Spec (21, 4,076); Training Language Models to Follow Instructions with Human Feedback (InstructGPT) (14, 3,379); Manhattan Institute — Measuring Political Preferences in AI Systems (Rozado) (13, 3,625); Measuring AI Ability to Complete Long Software Tasks (12, 2,772); METR — Measuring Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (12, 3,427); NERC 2025 Long-Term Reliability Assessment (LTRA) (11, 2,832); The OpenAI Files (11, 3,005); AI-designed viral genomes (7, 3,341); Misinformation on Misinformation: Conceptual and Methodological Challenges — Altay, Berriche, Acerbi (Social Media + Society, 2023) (6, 2,921).
  • Live-thread items named in the window with no page (score 3 each). government/army — Project Griffin, the Army's August 20 commercial solutions opening for agentic AI to process sensor data, part of the Army Rapid Defensive Cyber Systems program; the Army is named on 8 content pages and government/ holds air-force but not army, so this is the clearer structural gap of the group. models/gen-1-5 — named on 3 pages; deliberately left until the queued source is ingested, since the ingest note asks whether the model belongs on its own page or inside Generalist. Australia's Joint Select Committee on Artificial Intelligence, appointed August 20 with a report due November 30. Brazil's Bill 4356/25 on AI in the Unified Public Security System, approved in committee August 20. The Alabama Attorney General's investigation of OpenAI, which became public August 24 — held because the only source is a subscription lede.
  • Two live-thread items examined and declined, not deferred. Instinct / Spear Street Technology — the private-access assistant whose terms drew privacy objections between August 21 and 24. Zero prior mentions anywhere in the wiki, a company in private access, and every fact in the digest traces to individual testers' posts; a page written today would be a news item, not an encyclopedia entry. Ox Alpha — an unbranded model on OpenRouter whose developer is anonymous and whose leading fingerprint attribution weakened between the August 24 morning and evening digests. Its own analyst cautions that matching fingerprints prove shared infrastructure, not identity. Neither passes the quality gate; both are recorded here so they are not re-raised as oversights.
  • companies/tiktok — examined and declined. "TikTok" is named on 36 content pages, which reads mechanically as a large missing-company gap. It is not one: the mentions cluster on content moderation, minors' safety and platform speech law rather than on AI development, ByteDance is already covered at ByteDance / Doubao, and the enforcement thread now has a home at United States v. ByteDance Ltd. (COPPA). Same reasoning the 2026-08-23 run applied to companies/uber.
  • Carried unchanged from prior runs. Marvell (4 mentions, no company page, carried since 2026-08-21); models/deepseek-v3 (in-degree 49 on 608 words, routed to model-robustness-check on 2026-08-22 and still unaddressed); the Nvidia–Perplexity round (deferred 2026-08-24 — the substance has since been folded into Nvidia & TSMC — AI Compute Infrastructure and Perplexity, so this is close to closing); slice-11's source-robustness-check candidates led by California SB 53 — Transparency in Frontier AI Act; the White House release of 2026-07 from the primary-document baseline, instrument still unidentified after six runs.
  • The 28 date-stamped headers in lint-report.md (score 1). Down from 29 on 2026-08-24 — the reduction is real but the finding is unchanged in substance for nine cycles: the report still does not distinguish legitimate procedural histories (litigation pages, legislative timelines) from genuine timeline dumps, so the number cannot be driven to zero and is not a defect on its own. This is structural. It should be fixed in bin/lint-scan.py by excluding litigation/ procedural-history sections rather than re-listed each run.

Notes

  • The 2026-08-24 22:05 digest is still unprocessed (unprocessed_devlog=1). Two gaps actioned this run — Thomson Reuters and United States v. ByteDance Ltd. (COPPA) — draw on items that digest carries, so the nightly developments-log fold should update those pages rather than create new ones. Both were built from primary sources (the company's own disclosures; the DOJ press releases) rather than from the digest, and both were standing gaps before that digest existed: Thomson Reuters was named on 8 pages, and the COPPA action dates to 2024.
  • Primary-document scan (gap type 9) is clean in lane 1. Six digests scanned over three days: 2 cited primary-document URLs, 0 missing from Raw Sources/, 0 possible matches under another address. Lane 2 returned two keyword pre-filter hits, both read rather than actioned on the keyword: the DOJ/TikTok complaint, which became United States v. ByteDance Ltd. (COPPA), and the Thomson Reuters technical report, which the coverage says is "expected later" and therefore does not yet exist to capture.
  • Two dedupe drops, both correct. Anthropic's "Patterns and problems in emerging multiagent systems" (2026-08-13), updated in the 2026-08-23 08:11 digest — a Raw Sources/ file and a verification record already exist. Japan's draft generative-AI Principles Code, from the 2026-08-23 22:12 digest — already queued as INGEST-japan-generative-ai-principles-code.md.
  • No broken outbound links introduced. All wikilinks across the five new pages were resolved against the full page set after writing: zero unresolved.
  • Style scan clean. All five new pages checked programmatically against the banned-header, hyperbole, self-reference, standing-Predictions and date-header lists; zero hits. lint-scan.py after the run: banned_headers 0, self_refs 0, predictions 0, header_hype 0.
  • Broken-link count fell. 234 distinct / 257 refs on 2026-08-24 → 215 / 215, and the referenced-2+ lane went 23 → 0. The five pages created this run were not on the broken-target list, so they did not cause the drop; the author-entity backlog worked in earlier runs did.
  • Three unsourced assertions were caught and removed during drafting, recorded here rather than silently fixed: a founding year for Enfabrica written from memory before a source was found; a claim that AIA's chief executive is Eric Fanning, which the about page supports only as a name in headlines and not as a title; and a paragraph reading licence-and-hire transactions as reducing merger-review exposure, which no cited source states.
  • Confidence set honestly. All five new pages are confidence: medium. None met the three-independent-source bar.
  • Cap respected. Seven gaps actioned against a 6–10 cap; 34 candidates found. Six web pulls and four searches, no crawls.

One-line summary

Five pages created live — Thomson Reuters, Enfabrica, the DOJ/TikTok COPPA action, and the two defense trade associations filing against GSA's draft AI clause — plus two foundational sources verified and queued for review; the referenced-2+ broken-link lane reached zero for the first time in ten runs, and the slice-12 sweep produced ten source-robustness-check candidates that need a decision on whether that skill gets its own scheduled pass.

Categories: Briefings | briefing | gap-scan