Gap Scan — 2026-08-03

Scanned

Recent window: 4 New Developments Log/ files (2026-08-01 08:09 and 22:12; 2026-08-02 08:05 and 22:05) and 118 wiki pages edited in the last 48 hours. Rotation slice: 4 — entities/ A–G (173 pages).

Link scan over all mainspace pages returned 357 raw distinct broken [[wikilink]] targets; normalizing the trailing backslash reduces that to 275, leaving 9 targets with an inbound count of two or more distinct pages. The 82-target difference is the escaped-pipe artifact — targets written [[companies/anthropic\|Anthropic]] inside Obsidian table cells, captured as companies/anthropic\ by a regex that stops at | without stripping the escape. All nine survivors were already on the deferred backlog or curator-blocked, so no new alias error surfaced and no alias fix was made. This is the seventh consecutive run to report the same scanner defect.

Firecrawl was available. Six pulls were made — four scrapes and two searches — against a 6–10 cap; no bulk crawl. Four documents were read in full this run: OpenAI's August 1 publication, Gary Marcus's August 2 critique, the CAIS press release as reproduced by AIwire, and the news corroboration on the July 31 Washington preview. Two claims carried onto pages below come instead from the 2026-08-02 evening developments file, which recorded them from its own full-article scrape: Levent Alpöge's quoted reproduction claim and Thomas Bloom's assessment. Both are cited to the URLs that file read, and the Bloom quotation is marked on the page as reaching the record through an aggregator.

Gaps actioned (7 of 24 found)

New pages created (live)

• Astra — score 8 (live thread +3, dangling foundational reference +3, wiki core area +2), the run's highest. The 08-02 scan declined this gap on the quality gate, correctly: at that point Astra was a "tentatively named" family known only from three people briefed on unreleased plans, and a models/ page built on it would have been speculation. That basis no longer holds. On August 1, 2026 OpenAI used the name itself in a signed publication and put a substantial primary record behind it — ten enumerated results, a paper, per-solution reasoning walkthroughs, and Lean certificates in a public repository. The page records what the developer has published, what the pre-release reporting said and how thinly it is sourced, and the three independent challenges now on the record; it does not assert a shipped product. confidence: medium, sources_count: 8, release_date: unreleased, parameters and open_weights both undisclosed, safety_case explicitly recording that no system card or Preparedness Framework classification exists.

Two judgments worth recording. First, the page is written as a contested capability claim, matching how AI for Science already handles the May 2026 Erdős unit-distance result — OpenAI has not stated how many conjectures were attempted, so ten successes cannot be converted into a success rate. Second, an alias-resolution check was run against Google DeepMind's unrelated Project Astra, which is referenced on Google DeepMind and Healthcare — AI Deployment; the two are distinct products and the INGEST- task carries an instruction to keep them distinguishable after the fold.

Pages expanded (live)

• Center for AI Safety (CAIS) — 24 inbound references from 15 pages, 400 words, sources_count missing entirely, and untouched since 2026-06-06 despite a June 2, 2026 announcement that changed the organization's shape. Added a section on the appointment of Devin Kim as President — a newly created role; Dan Hendrycks remains Executive Director — and on the establishment of the Frontier Security Institute, a Washington-based organization under CAIS whose stated remit is to act as a translation layer between frontier AI developers and the Pentagon, intelligence community, Congress and allied institutions. FSI's senior staff are named with their prior service: Executive Director Isaac "Ike" Harris (23 years a U.S. Navy surface warfare officer, later policy adviser to the Secretary of Defense on China and technology security and a senior staff member of the House Select Committee on the CCP); COO Jeremy Pelter (former Acting Secretary of Commerce); Director of Research Aaron B. Frank (formerly RAND); VP of Communications Susan Malandrino. Also recorded: the two research lines the announcement names as in development, AI wellbeing and AI political manipulation. Added an ## Open questions section and five typed-relationship lines covering six targets. sources_count set to 6; last_updated → 2026-08-03; confidence held at medium. Original backed up to _meta/_revision-backups/entities/.

• AI for Science — the page's "AI-generated original results" section carried the May 2026 Erdős result and the July 21 Jacobian-conjecture claim but stopped short of the August 1 publication. Added a paragraph carrying the ten results' scope, the Lean formalization, the $2,000 token figure, and — in the same breath, per the section's existing register — the selection-denominator and full-cost objections and Alpöge's reported reproduction of five results with a shipped model. 2 new (Source: URL) cites; sources_count corrected 10 → 14.

Frontmatter and citation repairs (live)

• Dario Amodei — the wiki's most-linked entity page was missing a required field. 171 inbound references from 105 distinct mainspace pages — more than any other page in entities/ on either measure — and sources_count was absent from its frontmatter entirely while the body carried 17 distinct citations. The field is required by CLAUDE.md for all wiki pages. Set to 17, and a typed ## Relationships section — which the page had never had — added with twelve entries, including depends-on: for the risk-based-regulation and scaling-laws pages his positions rest on and litigates: for Anthropic v. United States (Pentagon ban challenge). last_updated deliberately not bumped: no sourced claim was added or changed.

• 81 of 173 slice-4 entities/ pages had understated or missing sources_count — 47% of the slice, against 44 of 81 (54%) measured on slice 3 yesterday and a comparable rate on slice 1. Recomputed from distinct (Source: URL) cites plus distinct wikilinks resolving into sources/, and corrected upward; no value was lowered. Five pages were missing the field outright: dario-amodei (→17), future-of-life-institute (→9), center-for-ai-safety (→5, later 6), american-innovators-network (→2), benjamin-guggenheim (→2). The largest understatements sit on high-reliance pages: alexandr-wang 7→20, americans-for-responsible-innovation 8→14, gary-marcus 6→13, g7 6→13, elon-musk 9→13, bureau-of-industry-and-security 7→11, controlai 3→10, erik-brynjolfsson 7→10, epoch-ai 6→10, anu-bradford 1→8, arvind-narayanan 3→9, eu-ai-office 4→8. Eight pages declared zero while carrying citations: cnil, fas, ai-now, anthropic-model-welfare, david-chalmers, david-ha, dpc, eliezer-yudkowsky.

last_updated was not bumped on the 81 repaired pages, on the precedent set 08-02: a count correction is not a content revision, and bumping the date would reset each page's confidence-decay window and mask real staleness.

A counting-rule note. This run counted wikilinks resolving into sources/ by basename as well as by the explicit [[sources/…]] form. Prior runs counted only the explicit form, which under-reports on entities/ pages, where the bare-slug style ([[cais-statement-on-ai-risk]]) is common. The stricter rule would have found 22 understatements here rather than 81; the difference is a measurement artifact, not a change in the wiki.

• OpenAI — the Astra paragraph still read "tentatively named" and described the family as known only from a single pre-release report, which stopped being accurate on August 1. Rewritten to preserve the July 31 reporting with its unnamed-sources caveat intact and to record OpenAI's own use of the name, the Marcus critique and the Alpöge reproduction claim — each with its own citation and its actor named — plus the link to the new page. 2 new (Source: URL) cites.

Recounting that page turned up a much larger error than the one being fixed: OpenAI declared sources_count: 17 while carrying 256 distinct citations, a fifteen-fold understatement on one of the wiki's largest pages. Corrected to 256. companies/ has not been audited for this defect — it is rotation slice 0, next due 2026-08-13 — and on the rates measured in concepts/ and entities/ it should be expected to carry the same problem at scale.

Queued — foundational sources

• OpenAI, "Ten advances in mathematics and theoretical computer science" (Publication, 2026-08-01) — score 8. The primary document behind a thread the wiki had been carrying entirely second-hand, and the anchoring citation for the page created live today. Verified: openai.com/index/ten-advances-in-mathematics/, the canonical host; HTTP 200 with sourceURL, og:url and url all matching the request exactly, so no redirect to a syndicated copy. Distinctive passages confirmed verbatim, including "an internal version of Astra, our next major model" and the arithmetic-formula lower bound of order n⁴/log n. Corroborated three ways: a full read of Gary Marcus's hostile August 2 critique, which links the same URL and independently reproduces the $2,000 figure and the "next major model family" wording from OpenAI staff posts; the independent scrape recorded in the 08-02 evening developments file, where every checkable item agrees; and India Today. Saved: Raw Sources/OpenAI - Ten Advances in Mathematics and Theoretical Computer Science (2026-08-01).md. Queued: INGEST-openai-ten-advances-mathematics-2026.md. Verification record: queue/gap-scan/proposed-sources/openai-ten-advances-mathematics-2026.md.

The 249-page paper and the reasoning-walkthrough PDF were not retrieved — both exceed the gap-scan retrieval budget — and the INGEST- task carries their URLs with an explicit instruction not to write proof-level claims from the announcement page, which enumerates results without arguments. Saving a fragment as though it were the source document is the failure source-fidelity was written against.

Authenticity-verification failures

None. Four documents were pulled and read this run — the OpenAI publication, Gary Marcus's critique, the CAIS press release as reproduced by AIwire, and the news corroboration on the July 31 Washington preview. All four resolved at HTTP 200 on hosts appropriate to their type with corroborating metadata. One provenance caveat is recorded rather than hidden: the CAIS announcement was read in the AIwire reproduction, which carries the tagline "Source: CAIS" and the full release text; safe.ai was not fetched directly, so the page cites the reproduction rather than the primary. A second, carried from the digest: Thomas Bloom's "big news" quotation rests on a search-result-only aggregator, and Astra says so where it appears.

Determined not to be gaps

• The data-center sales-tax repeal thread (New York's executive-order moratorium and Hochul's stated intent to repeal the sales-tax exemption; the Texas Senate Finance hearing of July 27 with 69 exemptions approved 2021–2025, 59 more certified in 2026, six of 20 audited projects out of compliance, and the ~$3.2 billion two-year revenue estimate; The Information's per-gigawatt cost calculation) — Data Center Siting / AI Power Politics carries 31 sources and already tracks the Virginia, Ohio, Texas and New York exemption fights, and New York Data-Center Moratorium (2026) carries EO 62. The new Texas figures are a fold for the developments-log cycle, not a missing page.
• Geoffrey Hinton's NCSL Legislative Summit remarks, Zvi Mowshowitz's August 2 assessment of the two lab containment incidents, and the Interconnects open-model review — Geoffrey Hinton, Zvi Mowshowitz and Nathan Lambert all exist and are the landing pages. Folds, not gaps.
• sources/…, wikilinks and page as broken targets — documentation prose that happens to sit inside double brackets. Eight, four and one reference respectively. Not gaps; unchanged from prior runs.

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

• Gary Marcus, "OpenAI's amazing — but vastly oversold — new model Astra" (2026-08-02) — read in full this run and not queued, to stay inside the cap. It sits on the foundational side of the borderline heuristic: it advances an original argument (the fallacy-of-composition frame, plus the verification-and-synthetic-data account of why mathematics does not generalize) rather than restating reporting, and it carries Ernie Davis's evaluation critique, which nothing else does. The strongest deferred item; score ~6.
• The Leiden declaration on AI and Mathematics (leidendeclaration.ai) — a primary text OpenAI cites by name in the publication queued today, with no page and no raw file. Score ~4.
• Stanford HAI / RegLab, "The Abundance of Reports and Incapacity of States" (Ho, Robitschek, Karthik, Malek, Ouyang; forthcoming, Yale Journal on Regulation) — surfaced by the 08-02 morning digest with substantial findings on AI-assisted statutory analysis across 50 states. A peer-reviewed-track paper, nearly always foundational. Score ~4.
• entities/cnil — in-degree 14, 710 words, and until today sources_count: 0 on a page with citations. The count is now correct at 2, but the page is a GDPR-enforcement anchor with two sources and no revision since 2026-06-06. Content expansion deferred; score ~4.
• The slice-4 2026-06-06 cohort. 35 entities/ A–G pages combine in-degree ≥8 with last_updated: 2026-06-06, the date of the v4.3 content/operational split — led by chris-olah (28), andrew-clearwater (26), gillian-hadfield (24), aisi-network (20), cosma-shalizi (19), deirdre-mulligan (17). Third consecutive slice to report this pattern (24 pages on slice 1, 8 on slice 2, 35 here). Across three slices this is plainly folder-wide migration residue rather than a set of individual gaps, and is better handled as a planned pass than as daily gap-scan residue — the third run to say so.
• New person pages implied by today's pulls, none created: devin-kim (CAIS President), isaac-harris and jeremy-pelter (FSI), levent-alpoge (the Anthropic mathematician whose reproduction claim is now load-bearing on two pages), matt-stoller (BIG, cited on the 08-02 digest with a detailed disclosure-reform proposal and no page anywhere in the wiki), ernie-davis, henry-yuen, thomas-bloom, noam-brown, lijie-chen.
• Carried unchanged, genuine but below the reference-count threshold: entities/samuel-weinbach and entities/ilhan-scheer (Aleph Alpha, in-degree 2 each); entities/clayton-christensen and entities/adam-smith (2 inbound references each, both pre-AI thinkers cited for a framework — still awaiting the curator ruling on whether the wiki wants biography pages for them at all); the long concept list carried since 07-31 (economic-possibilities-for-artificial-intelligence, algorithmic-decisionmaking, ai-and-language-models, talent-flow-china-us, data-broker-regulation, a-vision-of-democratic-ai, sortition, safety-training-methodologies, ai-coarse-grainings, cross-national-ai-policy-tracking, ai-history, biometric-identification, enterprise-ai, voice-ai); and concepts/inverse-cooking-problem / inverse-trust-problem, curator-blocked as coined terms since 07-10 and not re-actioned.
• [[_meta/briefings/weekly-2026-W19]] referenced from three mainspace pages (AI Macro-Prudential Policy, Nathan Lambert, Index). Seventh consecutive flag; still a schema question for the curator — whether mainspace may cite a briefing at all — rather than a mechanical rewrite.
• The bin/lint-scan.py escaped-pipe target-capture bug — seventh consecutive flag. Quantified again today at 82 of 357 targets, 23% of the reported count, and it puts five nonexistent targets at the top of the list. The one-line fix remains a .rstrip('\\') on the captured target, or the regex \[\[([^\]|#]+?)(?:\\?\|[^\]]*)?\]\]. Outside gap-identifier's remit.

Carried unchanged from prior runs

• The ingest queue backlog is cleared — this carry comes off the list. Every INGEST- task the 07-29 through 08-02 scans queued has been processed and archived: the two 07-29 tasks ("Pacing the Frontier"; the Hugging Face agent-intrusion timeline), the three 07-30 tasks (Weil, METR, Vending-Bench), the six 07-31 tasks plus the AI-2040 and pre-2024 backfill set, the five 08-01 tasks including Redwood, and the three 08-02 tasks (Beduschi, Suresh, the 2023 copyright suits). Today's OpenAI task is the only open item in Wiki/_meta/queue/. Prior runs carried a growing "still unprocessed" list that reached twelve on 08-01; it is now zero.
• 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), then California SB 53 — Transparency in Frontier AI Act (234), Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment) (178), Executive Order 14365 — Ensuring a National Policy Framework for AI (144), Clawed (105), California SB 243 — Companion Chatbots (98), NIST AI Risk Management Framework (AI RMF 1.0) (74).
• The high-confidence-on-sources_count: 1 question, unresolved across three sources/ slices and now visible in entities/ too: brookings-ai (10 references) and florence-gsell (9) both carried confidence: high on a declared single source before today's recount.
• 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); the entities/cdao / government/cdao duplicate — sharpened today, since cdao is among the most-referenced pages in this slice at 55 references from 33 pages, behind only dario-amodei, david-sacks (81), apollo-research and gary-marcus (58); models/mai-cyber-1-flash quality-gate decline (07-28); the Andon Labs entity_type and placement question (07-30); the Situational Awareness LP placement and entity_type-for-for-profit-non-developer question (08-02), which today's slice reinforces. Five slice-4 pages — constellation-energy, dominion-energy, nextera, softbank and sequoia-capital — carry an inline # comment inside the entity_type field explaining that nonprofit is a placeholder for a for-profit utility, holding company or venture firm. Whatever the schema decision, that comment is parsed as part of the YAML value and should be moved out of the field; a lint rule matching entity_type:.*# would catch all five.
• Model-slug inconsistency in models/: claude-opus-45, claude-opus-46, claude-opus-47 beside claude-opus-4-8. Fourth flag.
• The 14 dated dashboard-rebuild-failed notes from June, unrevisited for eight weeks; rebuilds have succeeded on every run since 07-31. They can be cleared on a curator's word.

Post-run verification pass

An independent audit of every touched file against house-style, source-fidelity and CLAUDE.md was run before this report was filed. It found real defects; all were repaired, and this report was corrected where it had misdescribed its own work.

• A false cross-reference, and the fact it pointed at was missing. Astra asserted that both of Alpöge's claims were "recorded on Claude Fable 5." Only the July 21 Jacobian claim was; the five-of-ten reproduction — load-bearing on two pages and declared as a contradicts: relationship — appeared nowhere on the page for the model that supposedly produced it. Claude Fable 5 now carries the claim in full, with the quotation, the named problems, the single-source caveat and the June 9 release date that gives it its force. The sentence on Astra was rewritten to be true.
• Two typed relationships misused on Dario Amodei. supports: was used in the political-endorsement sense for two state bills, where CLAUDE.md defines it as evidence for a claim; contradicts: was used for two people who criticised him, where it means evidence against a claim. Both recast as related:, and the litigation link moved to the defined litigates:.
• Five wiki self-references on the new page, including an entire meta-paragraph beginning "This page records what has been published…" and two instances of retrieval-process commentary ("the abstract was the only retrievable portion"; "rather than a source read in full") that describe the scan rather than the subject. All removed or rewritten. House style §2d bars this outright.
• An uncited, anonymized claim on OpenAI. "An Anthropic mathematician reported reproducing five of the ten" carried no source and left Alpöge unnamed while the name was in use one page over — the anonymization failure source-fidelity names. Now attributed, named and cited, with the Marcus critique given its own citation rather than folded into "endorsement and dispute."
• A figure capped in transit. Davis's estimate is "not less than $20,000 and possibly upward of $200,000." Astra recorded it correctly; AI for Science rendered it "$20,000 to $200,000," converting an open-ended floor into a bounded range. Corrected — a class-4 divergence committed one page away from its own correct version.
• A date hardened. The Jacobian claim is recorded on two existing pages as having "circulated on July 21" because the source is a newsletter; the new page restated it as "announced, on July 21." Restored to the hedged form.
• A URL dropped in transcription. The Lean-certificate link (github.com/openai/ten-proofs) was present in the fetched announcement but lost when the raw file was written, leaving a claim on Astra cited to a document that no longer contained it. Restored to the raw file body. Separately, "249-page paper" was stated as fact in the raw file's own metadata about a PDF that same metadata records as not fetched; it is now attributed to Marcus, who says he read it.
• open_weights: false on Astra asserted something no source states; changed to undisclosed. sources_count corrected 6 → 8 under the same counting rule the run applied to 81 other pages.
• An unattributed inference on Astra ("five of the ten results would be within reach of a model shipped two months earlier") replaced with the sourced release date plus an explicit note that the claim rests on one social-media post by an interested party and is unconfirmed.
• Report corrections. The prior quality-gate decline was attributed to the 08-01 scan; it was the 08-02 scan. The claim that "every claim added below rests on a document read this run" overstated enrichment — two claims come from the 08-02 developments file's own scrape, now said plainly. "In-degree 171" was an occurrence count, not a distinct-page count; every reliance figure in this report is now labelled as inbound references with the distinct-page count given where it matters.

Findings not acted on, referred onward: the ai-for-science tag used on Astra is not in any CLAUDE.md tag family, but it is established on AI for Science itself and is a schema question rather than a page defect. Neither Index nor _meta/retrieval-index.md was updated for the new page — a systemic gap, with 122 of 1,483 pages currently absent from index.md, and outside gap-identifier's remit.

One-line summary

One model page created for OpenAI's Astra now that the family is named on the record rather than only in pre-release reporting, two pages expanded and one corrected around it, the wiki's most-linked entity page repaired after being found to be missing a required frontmatter field, 81 slice-4 sources_count errors corrected, and the primary OpenAI publication verified, saved and queued for the curator's ingest.