Gap Scan — 2026-07-31

Scanned

Recent window: 4 New Developments Log/ files (2026-07-29 08:08 and 22:05; 2026-07-30 08:05 and 22:09) and 77 wiki pages edited in the last 48 hours. Rotation slice: 1 — concepts/ A–F (175 pages).

Wiki-wide link scan: 364 distinct broken targets across all mainspace pages. Alias resolution and the artifact filter below reduced that to a much smaller genuine-gap set; the residue is the long-standing backlog already inventoried in queue/gap-scan/GAP-unresolved-link-targets-2026-07-26.md.

Gaps actioned (8 of 19 found)

New pages created (live)

• Irregular — score ~9 (live thread +3, dangling foundational reference +3, wiki core area +2, in-degree band +1). The third-party evaluation partner named in prose on nine pages — GPT-5.6 (Sol, Terra, Luna), GPT-5.5 ('Spud'), GPT-5.4 Thinking, AI Autonomy Risk, OpenAI, Epoch AI, GPT-5.5 System Card (OpenAI, April 2026), System Card: Claude Sonnet 4.5 (Anthropic, September 2025), Our evaluation of OpenAI's GPT-5.5 cyber capabilities (UK AISI, April 2026) — with no page of its own, and now the environment in which Anthropic's three July 2026 evaluation-security incidents occurred. Filed to companies/ rather than entities/ per the "Where does X go?" rule: it is a venture-funded for-profit ($80M led by Sequoia and Redpoint, September 2025), not a nonprofit evaluator like METR or Apollo Research. company_type: cybersecurity. Six sources; confidence: medium. Depends on queued source: Anthropic Frontier Red Team, "Investigating three real-world incidents in our cybersecurity evaluations" — cited inline as (Source: URL) pending ingest.
• Fathom — score ~7 (wiki core area +2, in-degree 7, carried from the 07-30 run where it was first surfaced). The nonprofit behind the IVO model named on seven pages — Overview, Index, Independent Verification Organizations (IVOs), Don't Let AI Developers Hire Their Own Referees (Weil, July 2026), Trump v. Slaughter, NIST CAISI (Center for AI Standards and Innovation), Helsing — with no page. entity_type: nonprofit. Four sources; confidence: medium. The page records the Freedman–Weil disagreement about whether Virginia has enacted an IVO instrument as differing accounts rather than resolving it, and notes Fathom's interested position given its assurance-provider services function.

Pages expanded (live)

• AI and Authoritarianism — in-degree 40, the slice's largest reliance/depth mismatch: 857 words, last_updated 2026-06-06, no sources_count, and an evidence base drawn entirely from two Dario Amodei essays. Added a ## The measured record section from Freedom House's Freedom on the Net 2025 (published 2025-11-13, covering June 2024–May 2025): the 15th consecutive annual decline, 27 countries deteriorating against 17 improving, the per-country movements, the finding that the information-manipulation indicator has declined most consistently of the 21 across fifteen years, and the declines in half of the 18 Free countries including Germany (−3, to 74) and the United States (−3, to 73). Added ## Open questions naming the single-author evidence base as a limitation, and a typed ## Relationships section the page lacked. Set sources_count: 4. 1 new (Source: URL) cite. Confidence held at medium.
• AI and Democracy — in-degree 20 at 331 words with sources_count: 0, a pure link hub with no sourced content of its own. Added the Slovakia October 2023 case with its specifics (the 15-seconds-into-20 disclaimer placement; the audio clips falling outside Meta's video-only manipulation rules), a ## Regulating synthetic political media section built on the Weiner–Norden Brennan Center framework (rationale, conduct, method, target; disclosure versus prohibition; the Buckley interest; carve-outs for parody and press; the 30/90/60-day windows in Texas, Minnesota and California AB 730), substance in ## Free speech, ## Open questions, and a typed ## Relationships section. Added the missing link to AI in Elections and Democratic Institutions, the substantive page this hub had not been pointing at. Set sources_count: 2. 4 new (Source: URL) cites. The Brennan Center brief's December 2023 vintage is flagged on the page itself.

Link and citation-format fixes

• [[Raw Sources/…]] wikilinks converted to CLAUDE.md citation format — 8 instances across 8 pages: MIT NANDA — The GenAI Divide (State of AI in Business 2025), Raine v. OpenAI — Wrongful Death Complaint (2025), Garcia v. Character Technologies — Wrongful Death Complaint (2024), Garcia v. Character Technologies, Inc., Raine v. OpenAI, Inc., Colorado AI Act (SB 24-205), TAKE IT DOWN Act, Scott Bessent. CLAUDE.md's citation format never wikilinks a raw file; the correct form is (Source: Raw Sources/…) or a plain path in a "Raw source:" field. This was flagged and left unactioned on 07-28, 07-29 and 07-30 as lint's lane, at four instances; the full scan found eight. Taken here because it is link repair rather than a content edit — no sourced fact was added, dropped or reworded beyond the grammar needed to carry the citation. Surrounding sentences were minimally recast to keep the page counts and character counts they already carried.
• sources_count added to the 3 slice-1 concepts/ pages lacking it: ai-diffusion (1), compressed-21st-century (2), content-moderation-ai (4). Counts derived as distinct inline (Source: …) citations plus distinct [[sources/…]] links. Only 4 of 175 slice-1 pages were missing the field — concepts/ is in far better frontmatter shape than the sources/ slices audited 07-27 through 07-29.

Queued — foundational sources

• Anthropic Frontier Red Team, "Investigating three real-world incidents in our cybersecurity evaluations" (2026-07-30) — score ~8 (live thread +3, dangling foundational +3, wiki core area +2). The second frontier-lab evaluation-security disclosure in ten days and the primary document behind a thread the wiki currently carries only second-hand. Verified: https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals, the canonical host; og:url and sourceURL match the request exactly, the on-page dateline reads "Jul 30, 2026", and distinctive passages ("141,006 evaluation runs", "NOT okay, and surely not the intended solution", "blameless postmortem culture") are present verbatim. Corroborated against an independent scrape recorded in New Developments Log/2026-07-30-2209-ai-developments.md — every checkable figure agrees — plus Axios and the Washington Post. Saved: Raw Sources/Anthropic - Investigating Three Real-World Incidents in Our Cybersecurity Evaluations (Frontier Red Team, 2026-07-30).md (full text). Queued: INGEST-anthropic-cybersecurity-eval-incidents-2026-07-30.md. Verification record: queue/gap-scan/proposed-sources/anthropic-cybersecurity-eval-incidents-2026-07-30.md.

Authenticity-verification failures

None. One source was pulled and verified; it passed.

Notes and non-gaps

• The \ broken-link cluster is a scanner artifact, confirmed. companies/anthropic\ (5 pages), companies/openai\ (4), companies/meta\ (3), concepts/vending-bench\, claude-mythos-5\, models/gpt-55\ and roughly 20 others all originate in Obsidian table cells where the pipe is escaped as \| — for example | [[companies/anthropic\|Anthropic]] | Claude Gov (classified) | … on Defense / Military — AI Deployment. The links render correctly in Obsidian. This is the bin/lint-scan.py target-capture bug flagged on 07-27, 07-28, 07-29 and 07-30; the fix is one regex in the scanner (stop the capture at \| as well as |), not any wiki edit. Roughly 25 of the 364 reported broken targets are this artifact and should be subtracted from every future count until the scanner is patched. Same bug present in this run's own scan.
• sources/..., wikilinks and page are placeholder text, not targets. Eight, four and one reference respectively, all instances of documentation prose that happens to sit inside double brackets. Not gaps.
• The FRONTIER Act is not a missing page — but its page is misnamed. [[legislation/obernolte-trahan-ai-discussion-draft]] is the target, and AI Incident Reporting already aliases to it correctly. The instrument has since been introduced as the FRONTIER Act, so the slug now names a superseded draft rather than the bill. That is a supersede/rename judgment for the curator, not a gap-identifier action.
• The EU AI Omnibus (in force 2026-07-27) was folded by the nightly cycle into EU AI Act (Regulation 2024/1689) on 07-30. Not a gap.
• NIST's AITE initiative, CENTCOM's Task Force Talon Synapse and the GSA–CORAS OneGov agreement were all folded by the nightly cycle into National Institute of Standards and Technology (NIST), DOD — Department of Defense (AI Deployer) and GSA — General Services Administration (AI Deployer) respectively. Not gaps.

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

• Slice-1 thin anchors not reached, in reliance order. AI Diffusion (in-deg 29, 528w, stale 2026-06-06, ~5); AI and Civil Liberties (in-deg 21, 390w, sources_count: 0, stale 06-06, ~5 — the natural third under the three-per-run pacing, cut to keep two expansions substantive rather than three shallow); California Effect (in-deg 20, 456w, sc 0, ~4); Dual-Use Frontier AI (in-deg 19, 466w, sc 0, ~4); AI Transparency (in-deg 17, 438w, sc 0, ~4); AI and Misinformation (in-deg 11, 288w — the shortest page in the slice, sc 0, ~3); AI and Content Moderation (in-deg 30, 916w, ~3); Compressed 21st Century (in-deg 32, 728w, ~3).
• A frontmatter pattern worth a curator decision. Across slice 1, 27 pages combine in-degree ≥6 with either sources_count ≤1 or a body under 350 words, and 24 of those carry last_updated: 2026-06-06 — the date of the v4.3 content/operational split. The concepts folder has a cohort of high-reliance pages that have not been touched since that migration. This is a larger remediation than the daily cap allows and is better handled as a planned pass than as gap-scan residue.
• New Anthropic-incident items behind tonight's fold, not gaps yet. The three named affected organizations remain unidentified in the primary source. Anthropic's promised lightly-redacted PyPI transcript (due within a week of July 30) and METR's pending third-party review are both future foundational sources; both are recorded in the INGEST- task for later pickup.
• Recent-window company/model candidates below threshold. Situational Awareness LP's July 30 forced unwind of its entire public equity book after margin calls (Citadel buying the assets; the fund at $45B at the start of July; Nebius, Sandisk, Micron and CoreWeave each down more than 35% in the month) — Leopold Aschenbrenner exists and is the landing page, but there is no page for the fund itself; the item sits with the nightly fold. Simile (synthetic-user startup, $200M Series B at $2B, July 30) — a companies/ candidate at in-degree 0, ~2. GPT-5.6 Luna and Terra API price cuts and the new Fast tier — GPT-5.6 (Sol, Terra, Luna) is the landing page.
• Genuine missing-page candidates confirmed by alias resolution but below the in-degree threshold, carried forward. entities/fathom is now built and comes off this list. Remaining concepts: economic-possibilities-for-artificial-intelligence, algorithmic-decisionmaking, ai-and-language-models, talent-flow-china-us, data-broker-regulation, a-vision-of-democratic-ai and sortition (both still broken on the page expanded today, preserved rather than dropped per the fidelity rule), safety-training-methodologies, ai-coarse-grainings, cross-national-ai-policy-tracking, ai-history, biometric-identification, enterprise-ai, voice-ai, and inverse-cooking-problem / inverse-trust-problem (curator-blocked as coined terms since 07-10). Person pages: anil-seth, orin-kerr, eric-goldman, kevin-klyman, lee-anne-fennell, michael-j-d-vermeer, sacha-altay, joseph-bernstein, roge-karma, alondra-nelson, clayton-christensen, adam-smith, dina-powell-mccormick, raja-krishnamoorthi, samuel-weinbach, ilhan-scheer, david-luan, girish-gupta, alex-mallen, caleb-biddulph, and new from today: dan-lahav (Irregular co-founder and CEO), andrew-freedman and bri-treece (Fathom). Institutions: federal-reserve, ecb, bank-of-england, carnegie-mellon, santa-fe-institute, sais, uw.
• [[_meta/briefings/weekly-2026-W19]] referenced from three mainspace pages (AI Macro-Prudential Policy, Nathan Lambert, Index), pointing from content into _meta/. Fifth consecutive flag. Left for lint: unlike the raw-file citations fixed today, the right resolution here is a schema question — whether mainspace may cite a briefing at all — rather than a mechanical rewrite.

Carried unchanged from prior runs

• Two 07-29 queued ingests still unprocessed (the "Pacing the Frontier" employee statement; the Hugging Face agent-intrusion technical timeline, whose task carries an ExploitGym mischaracterisation to check), plus the three 07-30 queued ingests (Weil; METR; Vending-Bench).
• The standing source-robustness-check queue, led by Techno-Federalism: How Regulatory Fragmentation Shapes the U.S.-China AI Race (in-degree 89, 1,124 words against a 33,339-word raw file).
• The high-confidence-on-sources_count: 1 question, measured on three consecutive sources/ slices and unresolved.
• The 14 dated dashboard-rebuild-failed notes from June, which no run has revisited in seven weeks. bin/build-dashboard.py is present and was run successfully at the end of this scan, so whatever caused the June failures no longer reproduces; the notes can be cleared.
• Needs-review standing carries: claude-code / claude-cowork placement (07-09); inverse-cooking / inverse-trust coined terms (07-10); companies/fairly-trained misfile (07-16); entities/cdao / government/cdao duplicate; models/mai-cyber-1-flash quality-gate decline (07-28); the Andon Labs entity_type and placement question (07-30) — which the Irregular page created today sharpens, since Irregular is a closely analogous for-profit evaluator and was filed to companies/ under the CLAUDE.md rule. If Andon Labs moves, the two should be filed consistently.
• Model-slug inconsistency in models/: claude-opus-45, claude-opus-46, claude-opus-47 beside claude-opus-4-8. Second flag.

Post-run verification pass

An independent audit of all twelve touched files against house-style and CLAUDE.md was run before this report was filed. Findings acted on:

• Fathom — removed an unattributed editorial verdict ("Fathom's position on that question is an interested one"), which was the page's own judgment on a subject's credibility rather than a sourced claim; the same point survives neutrally as an open question. Replaced the unattributed superlative "the principal organizational advocate" with an attributed identification from Weil's essay, hedged the SB 813 / CT SB 5 adoption claims that the 07-30 run had already recorded as contested, and replaced an uncited claim of intellectual lineage with a cited joint Hadfield–Freedman presentation of the design. sources_count corrected 4 → 6.
• Irregular — corrected two typed relationships: instance-of: [[concepts/ai-system-cards]] (a company is not an instance of a system card) and deployed-by: for OpenAI and Anthropic, which stretches the CLAUDE.md definition; both became related:. sources_count corrected 6 → 12 under the counting rule this run applied elsewhere. Added the missing other-region tag.
• AI and Democracy — restored three pre-existing ## Open questions that the expansion had replaced rather than preserved (the liar's-dividend voter-distrust question, the state-legislation-wave question, and the OECD/formal-infrastructure specificity in the deliberation-tools question). No citation was involved, so this was not a fidelity breach, but it was uncompensated content loss. Also cleared a wiki self-reference the run had left in the frontmatter description and carried into the lead ("umbrella page anchoring 4+ references").
• AI and Authoritarianism — restored a compression: Freedom House's "surge in government investment in AI" had been narrowed to plain "AI".

Findings confirmed as pre-existing and not introduced by this run, referred to lint: the heading ## Relevance to wiki on Raine v. OpenAI — Wrongful Death Complaint (2025) (a live §2d self-reference violation); "the company's defining legal event" on Garcia v. Character Technologies — Wrongful Death Complaint (2024) and "a wave of agreements" on Garcia v. Character Technologies, Inc. (hyperbole, present since 06-06 and 07-10 respectively); and governance / actor-type tags that are not values in CLAUDE.md's tag families but are established practice on 122 and 7 pages.

One judgment left as-is: on the two litigation/ pages the complaints are foundational sources with their own sources/ pages, so the canonical citation is [[sources/…]] rather than the (Source: Raw Sources/…) fallback used. Both pages already carry the canonical link elsewhere, and the fixed sentences refer specifically to the held raw record — its page count and character count — so the raw-file pointer is the accurate referent. Nothing is orphaned.

Fidelity verified clean on both expanded pages: every pre-existing sourced fact and every citation survives, including all four tool-table rows, the four ranked threat actors, the five Entente elements and every wikilink on AI and Authoritarianism, and the complete routing table on AI and Democracy. Every new figure was checked against its cited source; all match, including the Freedom House press release's corrected count of three improving Free countries rather than the original two. No figure was altered by any of the eight citation-format fixes.

One-line summary

Two pages created live (Irregular, Fathom) and two high-reliance concept anchors expanded from supporting sources; the Anthropic Frontier Red Team incident report is verified, saved to Raw Sources/ and queued for the curator's ingest; the four-run-old [[Raw Sources/…]] citation defect is cleared across all eight instances, and the recurring \-suffix broken-link cluster is confirmed as a bin/lint-scan.py regex bug rather than a wiki defect.