Gap Scan — 2026-06-21

Scanned

Recent window: 4 dev-log files in last 48h (2026-06-20 05:05 / 15:09 / 22:06 and 2026-06-19 15:09 — Anthropic export-control fallout and ~$1T valuation/IPO, Z.ai GLM-5.2, John Jumper to Anthropic, Noam Shazeer/Dean Ball to OpenAI, Pew "Americans and AI 2026," OpenAI o3 rare-disease diagnoses, Anthropic state cyber program, Amazon/OpenAI biopic, Texas data-center siting, Norway school AI ban); ~30 wiki pages edited in last 48h. Rotation slice: 3 — concepts/ P–Z.

Gaps actioned (6 of ~14 surfaced)

New pages created (live)

• AI Control — broken [[concepts/ai-control]] (referenced from Scalable Oversight) plus prose use of "AI control" across the alignment cluster, no page under any name. Core safety/alignment area. Built from the founding Greenblatt–Shlegeris–Sachan–Roger paper (arXiv:2312.06942, ICML 2024): control framing, trusted/untrusted models, control evaluations (red-team/blue-team), and the protocol families (trusted monitoring, trusted editing, untrusted monitoring, defer-to-trusted). confidence: medium. Depends on queued source (INGEST-ai-control-intentional-subversion-2026-06-09.md, already in queue) — cited inline with a provenance note.
• Weak-to-Strong Generalization — dangling reference across 6 pages (AI Alignment, Automated Alignment Research, Scalable Oversight, Recursive Self-Improvement (RSI), Jan Leike, Jack Clark) with no page. Core superalignment concept. Built from the OpenAI Superalignment paper (arXiv:2312.09390): the weak-supervisor/strong-student analogy, the performance-gap-recovered metric, and the paper's own caveats. confidence: medium. Depends on queued source (INGEST-weak-to-strong-generalization-2026-06-21.md) — cited inline with a provenance note.

Pages expanded (live)

• Professional Services AI Adoption — thin anchor: 449 words, sources_count: 0 (no citations), in-degree 6. Added a sourced "Survey evidence" section (Thomson Reuters Future of Professionals 2025: enterprise GenAI adoption roughly doubled year over year, tax firms 8%→21%; AMA augmented-intelligence physician sentiment: physician AI use roughly doubled 2023→2024) with a caveat distinguishing self-reported adoption from measured substitution. sources_count 0 → 2; last_updated 06-06 → 06-21. [2 new (Source: URL) cites] Every prior fact and link preserved.

Link aliases fixed

• [[concepts/x-risk]] → resolved to AI Existential Risk in Longtermism (2 instances). "x-risk" is an abbreviation of existential risk; ai-existential-risk already appeared in the same prose sentence and the Relationships list, so the duplicate alias and its *(stub)* markers were removed rather than repointed.

Queued — foundational sources

• Americans and AI 2026: Chatbots, Smart Devices and Views on Impact (Pew Research Center, 17 June 2026) — foundational institutional report on the live public-opinion thread (Public Opinion on AI rests on three inline cites to it with no sources/ page; also named in the 06-20 dev-log). Verified: pewresearch.org (institution's own host; six ATP authors; 5,119 adults, Feb 17–23 2026; report PDF/topline/questionnaire present; figures match the dev-log). Saved: Raw Sources/Americans and AI 2026 - Chatbots, Smart Devices and Views on Impact (Pew Research Center, 17 June 2026).md. Queued: INGEST-pew-americans-and-ai-2026.md (+ verification record in proposed-sources/).
• Weak-to-Strong Generalization (OpenAI Superalignment team, Dec 2023) — foundational paper behind the new weak-to-strong-generalization page. Verified: arXiv:2312.09390 (primary host; corroborated by ar5iv and ACM DL). No raw file saved (arXiv PDF fetched fresh at ingest). Queued: INGEST-weak-to-strong-generalization-2026-06-21.md (+ verification record).

Authenticity-verification failures

• None this run. Both pulled/named sources verified against their canonical primary hosts.

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

• companies/crusoe broken-link count (8) — page exists; the Meta 1.6 GW Crusoe deal (06-18/06-19 dev-logs) folds into it via developments-log, not a gap. No action.
• Central-bank entity stubs (entities/federal-reserve, entities/ecb, entities/bank-of-england, in-degree 2 each) — better created as a cluster; deferred repeatedly (06-17 → 06-20). Score 2–3.
• entities/fhi / Future of Humanity Institute (in-degree 2) — blocked on the staged FHI-retrospective ingest deciding the landing page. Score 3.
• concepts/ai-persuasion — live thread but 0 wiki in-degree (topical hole, not a dangling link). Score 3.
• sources/anthropic-pwc-expanded-partnership (in-degree 5–6) — supporting-class partnership mis-linked as a sources/ page; lint territory (fold to inline (Source:)). Recurring. Score 2.
• concepts/foundation-models (in-degree 1–2) — likely folds into Frontier Models; lint territory. Score 1.
• claude-code (in-degree 8) — perennial; blocked on a products/ placement schema decision. Needs a call from the user. Score 3.
• entities/buck-shlegeris — newly relied on by the live concepts/ai-control page (Redwood Research cofounder); candidate for a future run. Score 2.
• concepts/a-vision-of-democratic-ai — a Raw Sources/ file exists but no sources/ page; link should point to sources/ once ingested, not a new concept page. Defer to ingest. Score 1.
• Norway near-ban on AI in elementary schools (06-19/06-20) — international AI-in-education regulation; folds into AI in Education via developments-log rather than a standalone legislation/ page. Score 2.

One-line summary

Created two live safety-concept pages (AI Control, Weak-to-Strong Generalization); expanded the zero-source Professional Services AI Adoption with two cited surveys; resolved the x-risk → ai-existential-risk alias in Longtermism; and queued two verified foundational sources (Pew "Americans and AI 2026"; OpenAI "Weak-to-Strong Generalization") for review. No verification failures. Three live page changes were committed; the dashboard rebuild was blocked by the recurring iCloud sync-duplicate folders (Wiki/_meta/site/* 2 — "Operation not permitted" on delete; the build's initial rmtree(SITE) fails). The graph is updated; the generated dashboard view is stale until a one-time Finder cleanup of the * 2 directories lets bin/build-dashboard.py run.