Gap Scan — 2026-06-20

Scanned

Recent window: 4 dev-log files in the last 48h (2026-06-18-1511, 2026-06-18-2205, 2026-06-19-1509, 2026-06-20-0505) plus ~40 wiki pages edited in the last 48h — the dominant live threads being the Commerce/BIS export-control action against Anthropic's Fable 5 / Mythos 5 (and Trump's June 19 walk-back), the OpenAI GPT-5.6 prep and the Shazeer / Dean Ball / Jumper talent moves, the Z.ai GLM-5.2 open-weights release, DeepSeek's $7.4B raise, the GSA AI-procurement clause, and the EA/alignment cluster (automated-alignment-research, ai-existential-risk). Rotation slice: 2 — concepts/ G–O.

The dominant live threads already have landing pages that the nightly developments-log cycle folded (anthropic, openai, deepseek-company, glm-5-2, gsa, dean-ball, howard-lutnick, no-fakes-act, eu-ai-act, etc.), so the budget went to the alignment-cluster gap the recent edits exposed, a thin G-O-slice anchor with zero sources, a clean link-alias fix, and two verified foundational sources from the live procurement and export-control threads.

Gaps actioned (5 of ~14 found)

New pages created (live)

• Scalable Oversight — dangling concept with in-degree 7 (automated-alignment-research, rlhf, geoffrey-irving, arc, and the source pages concrete-problems-ai-safety, instructgpt-paper, gpqa-benchmark) and no page under any name. Core safety/alignment area; surfaced via the recently edited automated-alignment-research (which links it as depends-on). Created as a confidence: medium page covering the problem statement and the approach family (RLHF, debate, reward modeling, iterated amplification, weak-to-strong generalization, ELK, sandwiching), with internal source links plus external (Source:) cites for the debate (arXiv:1805.00899), reward-modeling (arXiv:1811.07871), and weak-to-strong (OpenAI 2023) papers.

Pages expanded (live)

• Longtermism — G-O slice thin anchor: 490 words, in-degree 6, sources_count: 0 (no citations at all) despite anchoring the EA/x-risk cluster. Added a sourced canonical-works paragraph (Bostrom Superintelligence 2014; Ord The Precipice 2020 with the one-in-six estimate; MacAskill What We Owe the Future 2022; Greaves–MacAskill "strong longtermism"), the FHI closure (16 April 2024, Oxford), and a sourced critique paragraph (Torres-style critique; FTX/EA reputational shock). Folded 4 supporting (Source: URL) cites; sources_count 0 → 4. Also cleaned stale *(stub)* markers on links that now resolve to real pages (ai-existential-risk, effective-altruism, will-macaskill, toby-ord, nick-bostrom); kept *(stub)* only on still-missing x-risk and fhi.

Link aliases fixed

• [[concepts/ai-and-relationships]] → [[concepts/ai-companions]] (2 instances: sources/seemingly-conscious-ai-suleyman, sources/ais-delusional-spirals-hai). ai-companions is the existing parent concept for the relational-AI / emotional-dependence thread.

Queued — foundational sources

• GSAR LLM Data-Safeguarding Clause (FR Doc 2026-12205, 2026-06-17) — primary procurement-rule text the GSA-procurement thread (GSA — General Services Administration (AI Deployer)) currently rests on secondary reportage for. Verified: federalregister.gov (canonical .gov, document id resolves, date and distinctive content corroborated; full text ~87K chars). Queued URL-only (large doc, re-fetch fresh at ingest): INGEST-gsar-llm-data-safeguarding-2026.md (+ verification record in proposed-sources/).
• CSIS — "Commerce Restricted Access to Anthropic's Latest Models. What Comes Next?" (2026-06-16) — verified foundational source for the export-control live thread, standing in for the unavailable primary order (see failure below). Verified: csis.org (institution's own domain; named authors Koren/Kurland/Mehta; date and ECRA-2018 content corroborated by Just Security and Nextgov). Queued: INGEST-csis-commerce-anthropic-restriction-2026.md (+ verification record).

Authenticity-verification failures

• Primary Commerce/BIS export-control order against Anthropic — sought per the 2026-06-19 lint report's "suggested next sources." No public primary document exists: the operative instrument is a confidential, company-specific BIS "Is Informed" letter to Dario Amodei, not published in the Federal Register or on any trustworthy primary host. Rejected under the protocol (cannot be found on a trustworthy host); recorded in the CSIS verification record. The verified CSIS institutional analysis was queued in its place.

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

• entities/fhi / Future of Humanity Institute (in-degree 2) — deferred again per 2026-06-19: an FHI-retrospective source is already staged (INGEST-fhi-retrospective-asterisk-2026-06-09.md); revisit the entity page once that ingest decides the landing page. Score 3. (The expanded longtermism page now gives FHI more inbound context for that decision.)
• concepts/ai-persuasion(-capability) — live thread (Oxford persuasion study; lint flagged it) but still 0 wiki in-degree; a topical hole, not a dangling link. Revisit once it accrues inbound links or a foundational source is verified. Score 3.
• entities/eric-horvitz + entities/tom-mitchell — prominent AI scientists, but each referenced only from a single source page (scientific-progress-ai-horvitz-mitchell); below the high-reliance bar. Better created as a small cluster. Score 2.
• entities/pam-bondi (in-degree 2, both from government/pcast) — single referencing page; defer unless AG references recur. Score 1.
• concepts/foundation-models (in-degree 1, from frontier-models) — likely best folded into frontier-models rather than spun out; hand to lint. Score 1.
• sources/anthropic-pwc-expanded-partnership (in-degree 6) — recurring (06-16/06-17/06-19): supporting-class partnership mis-linked as a sources/ page; fold to inline (Source:) — lint territory, not gap-identifier. Score 2.
• Central-bank entity stubs (federal-reserve, ecb, bank-of-england, in-degree 2 each) — deferred 06-19; better as a cluster. Score 2.

One-line summary

Created concepts/scalable-oversight and expanded the zero-source concepts/longtermism; fixed the ai-and-relationships → ai-companions alias (2); queued two verified foundational sources (GSAR LLM data-safeguarding rule; CSIS Commerce/Anthropic analysis) for review; recorded one authenticity-verification failure (the non-public BIS "Is Informed" order). Two new live pages → dashboard rebuilt.