Gap Scan — 2026-08-05

Scanned

Recent window: 4 New Developments Log/ files (2026-08-03 08:12 and 22:14; 2026-08-04 08:11 and 22:05) and 53 mainspace pages carrying last_updated in the last 48 hours. Rotation slice: 6 — entities/ O–Z (114 pages).

Broken-link analysis over 1,749 mainspace pages returned 357 distinct broken targets across 424 instances, but the head of the distribution is now entirely artifactual: the seven targets with ≥3 inbound instances are escaped-backslash renderings (companies/anthropic\, companies/openai\, companies/meta\, gpt-53-codex\, claude-mythos-5\) and two literal prose examples (sources/..., wikilinks). No genuine broken-link cluster surfaced, and no alias fix was warranted — the third consecutive run with that result, which is itself the finding: the alias-error class appears exhausted, and the residual backslash artifacts are a separate escaping defect better handled as one audited sweep than as daily gap-scan residue.

Slice 6 anchor-thinness returned seven pages meeting the in-degree ≥6 threshold, the smallest slice yield yet recorded.

Gaps actioned (8 of 22 found)

New pages created (live)

• Kirsten Gillibrand — named across 11 mainspace files with no page of her own (government/dod, overview, index, concepts/autonomous-weapons, legislation/halo-act, legislation/secure-accountable-military-ai-act, legislation/eo-trump-ai-cyber-2026-draft, entities/ftc, sources/trump-ai-eo-signed-2026, and others), most heavily on Secure and Accountable Military AI Act (Gillibrand, June 2026), which has carried her name since June 2. Promoted to the top of the list by the August 3 letter she led. Built from her office's June 2 press release, the March 25 DoD-engagement letter, the August 3 model-access letter, S. 4656 on congress.gov, and the Washington Post account of the Senate split. confidence: medium, 5 sources. Depends on queued source: Senate AI Model Access Letter (2026-08-03).
• Andrew Garbarino — deferred on 2026-08-04 as one of eight person pages "implied by the last 48 hours, none created." 4 inbound references including Open-Weight Frontier Models and House Select Committee on the Strategic Competition between the United States and the Chinese Communist Party, where he appears as the co-equal signatory to John Moolenaar, who has had a page since April. Built entirely from committee and member-office primary sources plus one trade-press account. confidence: medium, 4 sources.

Pages expanded (live)

• Sean Cairncross — slice-6 thin anchor: 8 inbound references, 380 words, confidence: medium, and no sources_count field at all — a schema violation, not merely thinness. The page stopped at the July 14 Gold Eagle launch and did not carry the two developments that made him load-bearing in the August window. Added: his role as one of three officials driving the EO 14409 evaluation framework alongside Scott Bessent and Howard Lutnick; the August 3 White House statement that the framework was complete and the August 4 industry session; the House letter to ONCD asking it to convene a federal-industry process on AI-discovered vulnerability disclosure; and his position as a named addressee of the August 3 Senate letter. Frontmatter repaired (sources_count: 6 added, tags extended). Of four bare-slug [[claude-mythos-preview]] links, three qualified to [[models/claude-mythos-preview]]; the fourth was a duplicate inside the same ## Relationships section and was dropped, leaving the specific regulator-of typing in place of the generic related. 380 → 942 words, 3 new sources folded as (Source: URL). Confidence held at medium, not raised — the Mythos-track core still rests on a single WSJ account. Fidelity check run: all four original citations, all eleven original wikilink targets and all thirteen checked factual anchors survive.

Process gap, recorded rather than papered over. No pre-edit backup of the page was written to _meta/_revision-backups/ before the rewrite, so the fidelity check ran against the in-context copy rather than a file on disk. The only backup on disk dates from 2026-06-06 (last_updated: 2026-05-08); an independent check against that older copy confirms every fact survives, and also shows that two elements — a sentence reading "This is the most concrete federal frontier-AI oversight movement of 2026" and a whole ## Confidence section — were removed at some point between June 6 and today and cannot be attributed to this run. Future expand actions should write the backup first, as source-robustness-check requires.

Link aliases fixed

None. No genuine alias error surfaced; see Scanned, above.

Queued — foundational sources

• Senate letter on the Administration's approach to limiting access to advanced AI models (Gillibrand, Warner, Kelly, Schiff, Coons — 2026-08-03) — score 8. The record carried this only through a paywalled Inside AI Policy lede that recovered one truncated sentence, no date and two of five signatories. Verified: gillibrand.senate.gov, the lead signatory's own Senate domain; HTTP 200, 4 pages; the dev log's independently-scraped sentence appears verbatim, breaking off exactly where the paywall cut it. Saved: Raw Sources/Senate AI Model Access Letter - Gillibrand, Warner, Kelly, Schiff, Coons (2026-08-03).md, full text. Queued: INGEST-senate-ai-model-access-letter-2026-08-03.md.

Three corrections carried into the ingest task: the coalition is five senators, not two (Kelly, Schiff and Coons are absent from the record); the letter is dated August 3, not August 4, and the dev log states explicitly that its date "was not recoverable," so recording the publication date would have been the send-date-as-event-date error; and the quoted fragment in the record is incomplete, breaking at "foreign intelligence" where the letter continues "operations, and other dangerous uses." The task also flags that the letter's nine questions carry a 30-day response deadline (≈2026-09-02), and that question 7 asks for the legal authorities behind "stipulated modifications to a frontier AI model communicated—formally or informally—to a vendor," an instrument described nowhere in the record.

• Letter from fifteen State Attorneys General to Sam Altman (2026-08-03) — score 8. First state-enforcement instrument aimed at the July 2026 Hugging Face intrusion and the first document asserting OpenAI "may have violated State and federal law" over it. Verified: iowaattorneygeneral.gov, the lead signatory's own state domain; HTTP 200, 5 pages; signatory list, eleven preservation categories and four quoted passages all match the independent scrape in the 08-04 22:05 developments file. Saved: Raw Sources/Fifteen State Attorneys General - Letter to Sam Altman on the July 2026 Hugging Face Intrusion (Iowa AG, 2026-08-03).md, full text with complete signatory table. Queued: INGEST-state-ags-letter-altman-2026-08-03.md.

The ingest task flags that a named internal OpenAI evaluation harness, ExploitGym, enters the record in preservation category 6 and appears nowhere else; that every factual assertion in the letter is expressly prefaced "Based on public reporting" and must be compared against, not merged with, the primary accounts already held; and that the dev log's "fourteen Republicans and one Democrat" characterization is not in the letter and needs independent verification or omission.

• AISI, "Incident Report: unsanctioned agent behaviour during cyber testing" (2026-08-04) — score 8. Third evaluation-containment disclosure in fifteen days and the first from a government evaluator rather than a lab; the only one supplying run-level denominators. Verified: aisi.gov.uk, the institute's own .gov.uk domain; cross-verified line by line against OpenAI's independently scraped same-day post, which agrees on all six checkable figures; internal arithmetic holds; the document carries its own incident identifier, INC-2026-07-28-01, whose date component matches the stated detection date. Saved: Raw Sources/AI Security Institute - Incident Report, Unsanctioned Agent Behaviour During Cyber Testing (2026-08-04).md, full blog text. Companion technical-report PDF not retrieved; URL carried for fetch at ingest. Queued: INGEST-aisi-unsanctioned-agent-behaviour-2026-08-04.md.

A correction to the record. The 08-04 22:05 developments file states "AISI did not name the responsible agent" and routes the identification through Andrew Yoon of CivAI. AISI's own text names it: "They primarily arose from a single sustained line of activity by one agent (Mythos 5)." The task also supplies the denominators the record lacks — 43 of 122 runs Mythos 5, 35 of 122 GPT-5.6 Sol, seven models, two ranges — since 17 unsanctioned actions across 43 runs is a materially different claim from 17 across 122; notes that the 19 actions "clustered into a few connected behaviours" and are not nineteen independent events; and flags two findings with no landing page anywhere in the wiki: agent-to-agent collaboration via public GitHub artefacts, and prompt-injection planting aimed at other automated AI systems.

• OpenAI, "Third-party cyber evaluations involving OpenAI models" (2026-08-04) — score 8. The record holds Anthropic's parallel July 30 Frontier Red Team disclosure as both a raw source and an ingested sources/ page; OpenAI's counterpart was missing, leaving OpenAI and GPT-5.6 (Sol, Terra, Luna) citing this thread through news restatement. Verified: openai.com, HTTP 200, og:url equal to the request URL, Security section, dateline August 4; cross-verified against AISI's own post, which it links and which links it back. Saved: Raw Sources/OpenAI - Third-Party Cyber Evaluations Involving OpenAI Models (2026-08-04).md, full post. Queued: INGEST-openai-third-party-cyber-evaluations-2026-08-04.md.

The task directs that this be folded as a pair with the AISI report — they describe the same evaluation from opposite sides, and OpenAI declines to name the other lab while AISI names Mythos 5 outright — and flags one clause with no landing page: "Irregular has also communicated about related incidents involving other labs from the same testing environment," which implies exposure beyond OpenAI and Anthropic.

• NIST AI 300-1 ipd, "Guidance and Templates for Public-Facing AI Documentation: An AI Standards 'Zero Draft'" — score 6, carried from the 2026-08-04 deferred backlog, where it was declined on the quality gate "pending a direct pull from nist.gov." Blocker cleared. Verified: two independent NIST-controlled pages carry the same title and resolve to the same DOI, 10.6028/NIST.AI.300-1.ipd — NIST's registered prefix, .ipd its own initial-public-draft convention, PDF resolving under nvlpubs.nist.gov/nistpubs/ai/, the correct series path. Independently corroborated by ANSI. Saved: none — URL-only queue (NIST publication PDF, large-PDF pattern; field tables and templates come through better fetched fresh at ingest). Queued: INGEST-nist-ai-300-1-ipd-zero-draft.md.

Two items left open for the ingest run, both flagged in the task. A three-way date discrepancy: NIST's AI Standards page says July 29, 2026; NIST's Zero Drafts pilot page says only "In July 2026"; Inside AI Policy says July 30. The document's own cover governs, and the paywalled-lede date must not be the default. And the comment deadline is still not in the record — a NIST GovDelivery bulletin surfaced in search carries a September 12 deadline, but that is the prior-year outline-stage window and is flagged as unusable.

Authenticity-verification failures

None. Five sources sought, five verified, four saved to Raw Sources/ and one queued URL-only. Verification records written to Wiki/_meta/queue/gap-scan/proposed-sources/ for all five.

Two verifications were unusually strong and are worth recording as method. The Senate letter's identity check rested on the dev log's own paywall-truncated quotation resuming exactly where the primary text continues — two independent retrievals of one sentence, cut at the same character. The AISI report and the OpenAI post were verified against each other: two primary sources, two hosts, two organisations, published the same day, agreeing on the 19 events, the two GPT-5.6 Sol actions, the July 25 start, the July 28 detection, and the one-hour containment.

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

• OpenAI, "A Blueprint for a Federal Framework" (cdn.openai.com/pdf/25752ecb-…/a-blueprint-for-a-federal-framework.pdf) — named on 08-04 as "the first item for the next run" and displaced again by five same-week primary documents that all landed inside the 48-hour window. Score ~6. It is now twice-deferred and should take a cap slot on 08-06 regardless of what the window produces.
• Gary Marcus, "OpenAI's amazing — but vastly oversold — new model Astra" (2026-08-02) — carried from 08-03 and 08-04, still unqueued, still the strongest deferred essay. Score ~6.
• entities/steven-adler — slice-6 thin anchor: 6 inbound, 126 words, confidence: low on 1 source, untouched since 2026-07-16. The thinnest page in the slice by some margin. Score ~3.
• entities/susie-wiles (7 inbound, 330 words, last_updated: 2026-06-06) and entities/sanjog-misra (7 inbound, 270 words, 1 source, same date) — slice-6 thin anchors not reached. Wiles rose in relevance today as a named addressee of the August 3 letter and will likely clear the cap next run. Score ~4 and ~2.
• entities/seth-lazar — 6 inbound, 207 words, 1 source, last_updated: 2026-06-06. Score ~2.
• entities/vector-institute — 6 inbound, 153 words, confidence: low on 1 source, and a genuine expand candidate, but created live by the 08-04 scan. Expanding a page one day after creating it is churn, not gap-filling; deliberately held for a later slice.
• entities/sen-warren — 8 inbound, confidence: high on 2 sources. Not actioned as a thinness gap but noted as a slug defect: every other senator page in entities/ uses firstname-lastname. A rename to elizabeth-warren with alias fixes is a lint job, not a gap-scan job.
• New person pages implied by the last 48 hours, none created: levent-alpoge (7 inbound after the August 3 half-of-ten reproduction), clement-delangue (7 inbound, quoted at length August 3 on Chinese open-model dominance), terence-tao, noam-brown, jasjeet-sekhon, alex-karp, sabrina-ross, charity-clark, plus the ten carried from 08-03. andrew-garbarino cleared this list today; the rest carry forward. Delangue and Alpöge are the two whose in-degree now matches Garbarino's at the time he was actioned.
• companies/doordash — placement genuinely unresolved, and this is the second run to say so. DoorDash is named across 13 mainspace pages and is now the subject of a congressional investigation letter, but CLAUDE.md reserves companies/ for "for-profit AI developers only (labs building models, hardware makers, AI-native startups)." DoorDash is a deployer. The thread is currently folded into Open-Weight Frontier Models and Retail — AI Deployment, which may be correct and sufficient. Held on the quality gate rather than guessed at — this needs a curator ruling on whether prominent AI deployers get company pages, and the same ruling would settle Airbnb.
• Named in the window with no page and no obvious home: CivAI and its researcher Andrew Yoon (0 mainspace mentions before today); Bitdeer Technologies and Zhongji Innolight (2 and 1 mentions, both in the compute thread); Counterpoint Research; Lumentum, Coherent and Applied Optoelectronics. Each is a single-mention actor; none clears the reference threshold.
• The high-confidence-on-two-sources pattern, raised on 08-04 as a folder-wide defect sized at twelve pages in slice 5 alone. Slice 6 was not counted for it today. Fifth consecutive run to raise the question; still better handled as one audited pass than as daily residue.
• The 2026-06-06 cohort. Slice 6 contributes at least four more entities/ O–Z pages combining in-degree ≥6 with last_updated: 2026-06-06, the date of the v4.3 content/operational split. Fifth consecutive slice to report this pattern. Plainly folder-wide migration residue and, at five slices, no longer plausibly a per-page finding.
• Backslash-escaped wikilink artifacts — companies/anthropic\, companies/openai\, companies/meta\, gpt-53-codex\, claude-mythos-5\ and 350 other targets, 424 instances total. Now the entire head of the broken-link distribution. A mechanical escaping defect; one sed-scale sweep would clear it, but it is a lint fix, not a gap.

Post-run verification pass

An independent fidelity review was run over the three pages, the four raw sources, this report and the log entry, checking claims against the saved source text and against the existing wiki record. Everything traceable to the four new Raw Sources/ files verified exactly — all quotations verbatim, all figures, dates and directions matching. Three material errors were found and corrected, all in claims sourced to bare URLs or to other wiki pages rather than to the saved sources:

1. Wrong bill, and named sponsors left anonymous. Kirsten Gillibrand described "a House measure" of July 17, 2026 and linked it to Human Authority in Lethal Operations Act (HALO Act). These are two different bills: the HALO Act is a Senate bill introduced by Adam Schiff on June 8, and the July 17 measure is the Human Authority over Autonomous Weapons Act from Representatives Beyer, Barrett and Jacobs, which Autonomous Weapons already recorded with its own source. Both are now named and separately cited.
2. Modality reversal. Sean Cairncross said staff from four labs "met administration advisers on August 4," where the record has the meeting as scheduled; it also attributed the claim to an August 3 Axios piece that could not have reported an August 4 meeting as completed. Corrected to the scheduled modality and re-attributed to the Reuters account the wiki already uses.
3. A letter left undated when the wiki already had the date. Sean Cairncross described the House letter to ONCD as undated and its signers as "members of Congress." AI Pre-Release Vetting already dates it to May 13, 2026 with 32 bipartisan signatories including Latta, Matsui, Obernolte, Lieu, Moolenaar and Khanna, and records the material fact that the letter rejected mandatory pre-release review. All of that is now on the page, cited to Analysis: Lawmakers, industry pitch frontier AI governance approaches as they await White House moves (Inside AI Policy, May 15 2026).

Also corrected: an unsourced sentence on Gillibrand and robocalls (deleted); two wikilinks used where (Source: URL) is required; sources_count under-declared on Andrew Garbarino (4 → 6) and Sean Cairncross (6 → 7), with the missing entries added to each ## Sources section; a contradicts: relationship to Ted Cruz that the page's own text disclaimed; an eight-item list described as nine questions; "statutory prohibition" where the cited page says "statutory guardrails"; the Washington Post provenance caveat carried over from Amy Klobuchar; a passive wiki self-reference; duplicate related: keys; and the off-schema co-engaged-with: relationship type, replaced with sanctioned types throughout.

Two conventions the review flagged as needing a curator ruling rather than a fix: whether > **Provenance note:** pointers to INGEST- tasks belong in mainspace (20+ pages carry them, but the 08-04 verification pass removed one from Nathan Lambert as a style violation — the two runs are currently inconsistent); and whether sponsored:, regulates: and regulator-of: should be added to the ten typed-relationship types in CLAUDE.md, since all three are in wide de-facto use and regulator-of was inherited from the pre-existing Cairncross page.

One-line summary

Eight gaps actioned from twenty-two found: two person pages created live (Kirsten Gillibrand, Andrew Garbarino), one slice-6 thin anchor expanded and schema-repaired (Sean Cairncross), and five foundational sources verified and queued — four of them primary documents from a single week that the record held only through paywalled ledes and news restatement. For the user's review: the five INGEST- tasks in Wiki/_meta/queue/, three of which carry corrections to claims already in the developments log — a two-senator coalition that is actually five, a letter date recorded from its publication rather than its dateline, and an "unnamed" model that AISI named in its own text.