AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-07-17

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

Contents

Scanned

Recent window: 4 dev-log files (07-15 morning/evening, 07-16 morning/evening), 55 content pages edited in the last 48h (57 dated 07-15 to 07-17, minus the two pages this scan created). Rotation slice: 1 — concepts/ A–F (164 pages, thinness × in-degree crossed).

Tooling note: the firecrawl family was unavailable this session; verification used the built-in web search/fetch fallback the web-research skill sanctions for this case. All authenticity-protocol steps were still run per source.

Gaps actioned (8 of ~27 found)

New pages created (live)

  • Digital Markets Act (EU) (score ~7) — type-3 missing legislation page on a live thread: the 07-16 evening dev-log's lead EU item (two binding specification decisions ordering Google to open Android to rival AI assistants and share search data with AI chatbots) folded onto EU AI Act (Regulation 2024/1689) and four company pages, bringing to eight the mainspace pages naming the Digital Markets Act with no page behind it and no wikilink ever attempted. Alias check confirmed no page covers it under any slug. Built from the Commission's own DMA site (July 16 news article, fetched in full), the EC presscorner release, EUR-Lex (CELEX 32022R1925), the EC's April 23, 2025 Apple/Meta non-compliance decision page, and the already-cited Reuters/Bloomberg/Axios/CNBC coverage relocated from the referencing pages. First-mention backlinks added on all eight referencing pages (EU AI Act (Regulation 2024/1689), Apple, Meta AI, Microsoft, Amazon, Agent Supply Archetypes (BYO/Bowling-Shoe × Horizontal/Vertical), two Farahany source pages) and an entry added to Index. confidence: medium. [Depends on deferred source: the regulation's primary text — provenance note left on the page; see backlog.]
  • Guidelight Control Standard (v1.0) (score ~7) — type-3 missing standards page on a live thread, the 07-16 backlog's "possible standards/ page + source queue" item now matured: the nightly fold created Guidelight AI Standards but the standard itself — a formal, assessable standard, squarely standards/ territory — had no page while being named on AI Control and both entity pages. Built from the standards body's own catalog (guidelight.ai/standards, fetched; six principles confirmed verbatim) and Adler's July 15 explanation post; documents the companion Capability Testing and Transparency v1.0 standards and the self-administered-vs-third-party-assessable distinction. Added to Index. confidence: medium. [Depends on queued source: the Control standard full text.]

Pages expanded (live)

  • AI Safety Cases and Frameworks — the slice's most-relied thin anchor: in-degree 166 (highest in concepts/ A–F) at 873 words, sources_count missing, stale since 06-06 while the framework landscape moved. Added the RSP v3.4 deltas (effective July 8, 2026; from the verified anthropic.com host, mirroring Responsible Scaling Policy (RSP)), a new "Third-party assessment and standards bodies" layer (FLI AI Safety Index Summer 2026 grades — Anthropic C+, OpenAI C, Google DeepMind C, Meta D+, xAI F — from the verified futureoflife.org host; Guidelight's three assessable standards), and the July 14 Hassabis FINRA-modeled standards-body proposal at the institutional layer (cross-linked to AI Pre-Release Vetting). Frontmatter repaired (sources_count: 13). Every prior fact and citation preserved; backup in _meta/_revision-backups/. [4 new (Source: URL) cites]
  • AI Software Progress — slice thin anchor #2: 421 words at in-degree 46 on a single source (The Least Understood Driver of AI Progress), sources_count missing, stale since 06-06. Added Epoch AI's independent estimate (Ho et al. 2024, arXiv 2403.05812: compute for a given performance level halves roughly every 8 months, 95% CI 5–14 months — roughly 2–3×/year against the essay's 10×/year central estimate, giving the page a second anchored figure), the METR time-horizon cross-reference (Measuring AI Ability to Complete Long Software Tasks, ~7-month doubling), and a typed ## Relationships section. Frontmatter repaired (sources_count: 3). Backup in _meta/_revision-backups/. [2 new (Source: URL) cites]
  • Guidelight AI Standards (correction + expansion, in passing) — the page credited founding to Steven Adler alone; the founders' own announcement gives co-founders Page Hedley and Steven Adler (Adler as Chief Scientist), and the catalog lists three standards, not one. Corrected and expanded from the primary hosts; wired to the new standards page. sources_count 1 → 3; confidence stays low (organizational details still rest on the org's and founders' own publications). Backup in _meta/_revision-backups/.

Queued — foundational sources

  • OpenAI Global Affairs, "Mind the US-China Safety Gap" (July 15, 2026) — type-1 dangling foundational reference on a live thread (07-16 backlog carry, matured): named by title on US-China AI Competition: Different Races, Different Metrics, OpenAI, and Index with supporting cites only. Advances OpenAI's own five-Chinese-models red-team findings, the Concordia AI transparency survey recitation, and the same-day Public Policy Agenda launch. Verified: OpenAI Global Affairs' own Substack (metadata dated 07-15); Semafor same-day coverage; outbound identifiers resolve. Full raw saved: Raw Sources/Mind the US-China Safety Gap.md. Queued: INGEST-openai-mind-us-china-safety-gap-2026-07-17.md.
  • OpenAI, "GPT-Red: Unlocking Self-Improvement for Robustness" (July 15, 2026) — type-1 dangling foundational reference on a live thread: named on Overview, Jailbreaking and Red Teaming, OpenAI, and Index with supporting cites only. Self-play automated red-teaming at largest-post-training-run compute scale; 84%-vs-13% held-out prompt-injection results; will not be released. Verified: openai.com (exact title); corroborated by MIT Technology Review and MarkTechPost. Queued URL-only: INGEST-openai-gpt-red-2026-07-17.md.
  • Guidelight AI Standards, Control standard v1.0 — the primary text behind the new standards page. Verified: guidelight.ai/control on the standards body's own catalog (six principles confirmed verbatim on the catalog page); founding corroborated via Adler's announcement and LinkedIn. Queued URL-only: INGEST-guidelight-control-standard-2026-07-17.md.

Authenticity-verification failures

  • None. (Checks that did not become pulls: Gemini 3.5 Pro — the July 17 target date remains leak-based with no official Google announcement, no model card, and no API listing as of this morning; not paged, carried. H.R. 9619 text — congress.gov returned an empty page body to the fallback fetcher this run; last verified "text has not been received" 07-16; carried.)

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

  • DMA primary text (Regulation (EU) 2022/1925, EUR-Lex) — ingest queue candidate now that Digital Markets Act (EU) exists with a provenance note; large primary text, URL-only pattern. Score ~4.
  • Former Meta Employees v. Meta complaint — the litigation page (created 07-16) has no docket number; needs CourtListener identification before the complaint can be verified and queued. Score ~4.
  • models/gemini-3-5 — create once 3.5 Pro actually ships (target was today, July 17; unconfirmed this morning). Xi's WAIC keynote also July 17 — tonight's dev-log will carry it; check both next run. Score ~3.
  • Slice-1 residue (two-per-run pacing): AI Race Dynamics (in-deg 109, sources_count missing), AI and Authoritarianism (58, sc 0), AI and Productivity (57, 3,490w — frontmatter repair only), AI Governance (umbrella) (55, high with sc 0 — frontmatter), AI Diffusion (36, 528w), AI and Democracy (25, 406w), AI Consciousness (14, low/1). Score ~3–4 each.
  • US AI Regulatory Approaches Compared (in-deg 36) and Labor Disruption Timelines: Who Predicts What and Why (in-deg 21) — carried since 07-12; per the 07-16 note these belong to a dedicated crystallize/query session, not gap-scan's lanes.
  • Orphan-check sources-layer backlog (07-15): six state chatbot-act primary texts + the Model Spec blog post. Score ~2 each; batch decision when the queue drains.
  • Standing carries: claude-code/claude-cowork placement (curator-blocked), inverse-cooking/inverse-trust coined terms (needs-review 07-10), duplicate entities/cdao/government/cdao (lint), companies/reflection/reflection-ai duplicate (needs-review 07-15), companies/fairly-trained misfiling (needs-review 07-16), concepts/foundation-models disambiguation, concepts/safety-training-methodologies and concepts/a-vision-of-democratic-ai (in-deg 2), person/org pages at in-degree 2 (pam-bondi, masahiro-mori, anil-seth, eric-horvitz, orin-kerr, alondra-nelson, palmer-luckey, tom-mitchell, soufan-center, richard-dawkins, adam-smith, clayton-christensen), Helix (Figure AI Vision-Language-Action model) (next models slice). Score ~2–3.

One-line summary

Two live pages created (the eight-page dangling Digital Markets Act reference finally has a home wired from all referencing pages; the Guidelight Control standard moved from entity prose to a proper standards/ page), the slice's two most-relied thin concept anchors expanded (ai-safety-frameworks at in-degree 166; ai-software-progress gains an independent second estimate), a founding-attribution error on the Guidelight entity page corrected, and three foundational sources verified and queued for review (OpenAI's US-China safety-gap report with full raw saved, GPT-Red, and the Control standard text).

Categories: Briefings | briefing | gap-scan