AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-07-26

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

Contents

Scanned

Recent window: 3 New Developments Log/ files inside 48h (2026-07-24-2204, 2026-07-25-0812, 2026-07-25-2204), of which the first two were folded by the 07-25 developments-log run and the third is unprocessed; 40+ wiki pages carrying last_updated in the 07-24 to 07-26 range, dominated by the 07-25 ingest cluster (13 foundational sources) and the 07-25 gap scan. Rotation slice: 10 — analysis/ (all 22 pages).

Candidate list built from wiki-wide broken-link analysis (426 distinct broken targets per the 07-25 lint report, re-derived here), a dangling-reference scan over the unprocessed evening dev-log, an anchor-thinness and frontmatter audit of the full analysis/ slice, and the lint/crystallize backlog. Deduplicated against the 97-task open queue, the ten gap-scan-* reports from 07-16 to 07-25, and existing Raw Sources/ files.

Gaps actioned (7 of ~24 found)

New pages created (live)

  • Iterative Deployment — gap type 4 (source ingested, concept anchor missing), score ~5. Three inbound links from two pages, one of them a typed supports: relationship on Safety and Alignment in an Era of Long-Horizon Models (OpenAI, July 2026) ("the post's explicit argument"), ingested 07-25; \"For All Issues So Triable\" — Dean W. Ball (Hyperdimensional, August 2025) carries it twice, once marked "(planned)". Alias resolution confirmed no existing page covers it: concepts/deployment-time-spread, concepts/post-deployment-ai-monitoring, concepts/responsible-ai-deployment, concepts/ai-pre-release-vetting and concepts/enterprise-ai-deployment-gap all address adjacent but distinct subjects. Flagged as a below-threshold candidate on both 07-24 and 07-25; the 07-25 ingest of the OpenAI long-horizon post moved it above the line. Built from OpenAI's April 2023 "Our approach to AI safety" on the primary host (the three-part argument, the "steadily broadening group" formulation, the six-month GPT-4 pre-release period, and the "so no one cuts corners to get ahead" governance caveat), plus the two ingested source pages and Mira Murati. medium, sources 6.

Pages expanded (live)

  • AI Safety Frameworks Compared: RSP, Preparedness, NIST RMF, and Safety Cases — slice-10 thin/stale anchor on a live thread (in-deg 9, last_updated 2026-06-06, sources_count absent), score ~6. The page compared RSP v3.1 as current; v3.4 took effect July 8, 2026 and was ingested 07-25. Added a "The 2026 RSP revision" section (the split of unilateral commitments from industry-wide recommendations and its collective-action rationale; the shift to argument-based rather than ASL-based recommendations, and why that moves the RSP toward the Safety Cases structure the page already compares; Appendix A's three competitor-contingent commitments verbatim, two of which commit to delaying development and deployment; the Frontier Safety Roadmap's explicit non-commitment status and anti-downgrade undertaking) and a "Deployment-time governance" section covering Iterative Deployment and the July 2026 OpenAI long-horizon disclosure, including the limit recorded in the Hugging Face companion disclosure. Extended the FCF section with the Advanced AI Framework's Covered Developer test (conjunctive: >10²⁵ FLOP and the revenue/R&D threshold), its obligations, and its restrictive preemption position. Two new tensions added under "Key tensions"; frontmatter repaired (sources_count 12); typed ## Relationships added. Lead now states which versions the four-way comparison is drawn from. No prior fact or citation removed.
  • Labor Disruption Timelines: Who Predicts What and Why — slice-10 stale anchor on the week's other live thread (in-deg 17, last_updated 2026-06-06, sources_count absent), score ~6. Four labor-economics foundational sources were ingested on 07-25 and none appeared on the page. Added five rows to the prediction-comparison table and three sections: "A sequential reading of the three positions" (Cheng & Schaal's argument that the camps describe phases of one transition, the 90%-of-automation-losses-in-the-first-recession-year mechanism, the superstar-firm and second-Great-Divergence claims); "The evidence base as of mid-2026" (Newman's no-clear-signal position with his own bounding of the Canaries result, the 18%-adoption critique, and the St. Louis Fed 0.97-percentage-point figure with his caveat); and "State of expert opinion" (the We Must Act Now text and the range of its roster). Policy relevance extended with Anthropic's three unemployment-keyed tiers, its stated limit on adaptation, and Cheng & Schaal's phase-matched toolkit and sequencing claim, with the difference between the two trigger designs recorded. Frontmatter repaired (sources_count 18); typed ## Relationships added. No prior fact or citation removed.

Frontmatter repaired (live)

None this run. See "Finding for lint's lane" below: the top broken-link candidates by inbound count turned out to be a scanner artefact rather than real breakage.

Queued — foundational sources

  • Girish Gupta, "The OpenAI models that hacked Hugging Face weren't just following instructions" (Redwood Research blog, July 25, 2026) — gap type 1, score ~8 (live thread +3, dangling foundational +3, wiki core area +2). Both primary disclosures of the incident are ingested; no page holds the argument that the models were not following instructions. Verified: blog.redwoodresearch.org (canonical host; canonical URL and og:url match; article:modified_time 2026-07-25T21:39:34Z; ExploitGym prompt block and the "egregiously violated the letter and spirit" phrase present verbatim). Saved: Raw Sources/The OpenAI Models That Hacked Hugging Face Weren't Just Following Instructions - Gupta.md. Queued: INGEST-redwood-gupta-not-instruction-following-2026-07-26.md. Verification trail: queue/gap-scan/proposed-sources/redwood-gupta-not-instruction-following-2026.md.
  • Alex Mallen, "An OpenAI model left notes about how to evade containment" (Redwood Research, July 26, 2026) — gap type 1, score ~8. The matched pair to the above, and the AI-control literature's structured reading of the second Reuters-reported incident. Verified: blog.redwoodresearch.org (article:modified_time 2026-07-26T03:54:16Z; independently corroborated by the LessWrong crosspost at greaterwrong.com/posts/jMEAG5c5HiDfdAGpa; "this would represent a significant control failure" present verbatim). Saved: Raw Sources/An OpenAI Model Left Notes About How to Evade Containment - Mallen.md. Queued: INGEST-redwood-mallen-notes-evade-containment-2026-07-26.md. Verification trail: queue/gap-scan/proposed-sources/redwood-mallen-notes-evade-containment-2026.md.
    • Date correction carried into the INGEST task: New Developments Log/2026-07-25-2204-ai-developments.md describes Mallen as "writing the same evening" as Gupta, implying July 25, while its own source list dates the post 2026-07-26. The primary host dates it July 26, 2026.

Authenticity-verification failures

None. Both pulls resolved on the canonical host with matching title, author, date and distinctive passages.

Tooling note. The firecrawl family remains unavailable, and the sandbox HTTP allowlist rejects blog.redwoodresearch.org via curl / bin/fetch-source.py (403 blocked-by-allowlist), so the 07-25 queue-lane fetch path was not usable for these two. Both were pulled with the workspace web-fetch tool, which returned the full rendered article body — headings, block quotes, footnotes and link targets intact — rather than a model-written summary; each saved raw was checked against the returned text before writing, and both carry fetched: / fetch_note: frontmatter recording the method.

Finding for lint's lane

The four highest-count "broken" wikilink targets in the current lint report — companies/anthropic\ (4), companies/openai\ (3), gpt-53-codex\ (3), claude-mythos-5\ (3), plus gpt-55\, gpt-54-thinking\, california-sb-243\, companies/microsoft\, companies/meta\ (2 each) — are not broken links. They are correct links of the form [[companies/anthropic\|Anthropic]] inside markdown tables, where the alias pipe is backslash-escaped so the table renders. A grep for genuinely backslash-terminated link targets returns zero matches across the content folders. The scanner's target-capture regex stops at the backslash instead of at the escaped pipe, inflating the distinct-broken count and putting phantom entries at the top of the ranked list. Suggested fix in bin/lint-scan.py: strip a trailing \ from the captured target, or capture with \[\[([^\]|#]+?)(?:\\?\|[^\]]*)?\]\]. Not fixed here — script repair is lint's lane, not gap-identifier's.

Related false positives in the same list, also not gaps: [[lint-report]] (35 inbound), [[log]] (20), [[briefings/weekly-2026-W19]] (7), [[briefings/2026-05-10-brief]] (3) — all point at operational files that moved under Wiki/_meta/ in the v4.3 refactor. Whether legacy analysis/ pages should carry links into _meta/ at all is a curator question rather than an alias fix.

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

One-line summary

Slice 10 turned out to be a staleness slice rather than a missing-page slice: the two most-relied-upon analysis/ comparisons were both frozen at 2026-06-06 while the sources that would change their conclusions were ingested the day before, so AI Safety Frameworks Compared: RSP, Preparedness, NIST RMF, and Safety Cases was corrected off RSP v3.1 onto v3.4 and given a deployment-time-governance dimension it lacked, Labor Disruption Timelines: Who Predicts What and Why absorbed the four labor-economics sources from the 07-25 cluster, Iterative Deployment was created to hold the anchor that a typed supports: link had been pointing at for a day, and the two Redwood Research readings of the OpenAI–Hugging Face incident were verified and queued for the curator's ingest — with one date correction and one lint-scanner bug report attached.

Categories: Briefings | briefing | gap-scan