AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-08-04

Daily gap hunt — what the wiki is missing and how each gap was triaged.

Contents

Scanned

Recent window: 4 New Developments Log/ files (2026-08-02 08:05 and 22:05; 2026-08-03 08:12 and 22:14) and 182 wiki pages edited in the last 48 hours. Rotation slice: 5 — entities/ H–N (120 pages).

Link scan over 1,737 mainspace pages returned 357 raw distinct broken [[wikilink]] targets; normalizing the trailing backslash and dropping documentation placeholders reduces that to 269, leaving 30 targets with an inbound count of two or more. All 30 were already on the deferred backlog or curator-blocked, so no new alias error surfaced and no alias fix was made — the second consecutive run to find none. This is the eighth consecutive run to report the escaped-pipe scanner defect that produces the 88-target difference.

Firecrawl was available. Six pulls were made against a 6–10 cap — four scrapes and two searches; no bulk crawl. Four documents were read in full this run: the arXiv abstract page for 2606.03811, OpenAI Global Affairs' August 3 post, Klobuchar's November 2023 Senate release, and Heim's own publication list. A fifth, the RAND Perspective PEA3776-1 landing page, was read through its search-result body rather than a direct scrape, and is cited for the author biography and the report's own framing only. Two claims on the Heim page rest on search-returned excerpts rather than full reads and are marked below.

Gaps actioned (7 of 26 found)

New pages created (live)

The page's substance comes from a primary document the record did not previously carry: the November 15, 2023 Senate release for the AI Research, Innovation, and Accountability Act, read in full from klobuchar.senate.gov. That bill's two-tier structure — "critical-impact" systems in critical infrastructure, criminal justice and biometric identification facing Commerce testing standards and pre-deployment risk assessments; "high-impact" systems in housing, employment, credit, education, healthcare and insurance facing transparency reports — is the earliest form of the risk tiering that recurs in later federal proposals, and it was already a Klobuchar–Thune bill in 2023. Both senators' framings are quoted against each other, since they describe the same text in opposite registers ("common sense safeguards" against "limit government intervention"). confidence: medium, sources_count: 5.

Two judgments recorded. The 2026 bill section is written as reported and unconfirmed throughout: the Washington Post account rests on two people close to the discussions and an anonymous industry representative, three of the four offices did not comment, and no text exists. And the page states plainly that whether the bill preempts state law was not disclosed — the question on which the whole federalism thread turns.

Pages expanded (live)

  • Lennart Heim — the strongest genuine thin anchor in slice 5: 22 inbound references from 12 pages, 473 words, untouched since 2026-06-06, and — the defect that moved it up the list — confidence: high on sources_count: 2, below the three-source floor CLAUDE.md sets for a high rating. Expanded to 1,453 words from his own publication list and the RAND Perspective landing page. Four changes:
    • Affiliation corrected. The page described him as "a policy researcher at the RAND Corporation." RAND's own author note gives the title: associate information scientist at RAND and professor of policy analysis at the Pardee RAND Graduate School, leading compute research in the Technology and Security Policy Center within RAND Global and Emerging Risks. Leaving a named role generic is one of the four failure classes in source-fidelity.
    • Publication record built out, organized in four clusters rather than as a list: compute as a governance instrument (the 2024 GovAI paper on which he is second author behind Girish Sastry, the cloud-intermediary and know-your-customer papers, the compute-thresholds paper with Koessler); export controls and the chip supply chain (the sole-authored RAND Perspective PEA3776-1 on the AI diffusion rule, the hardware-enabled-governance working paper, the IaaS submission, the UAE report); compute measurement and trends (the Epoch series, Compute at Scale, the compute-efficiency and compute-divide papers); and technical AI governance more broadly. A new Regulatory submissions section records six US and UK filings, including the January 2023 comment on the BIS October 7 rule.
    • Positions section extended with his own framing of the enforcement problem — moving "the unit of governance from AI chips, which are hard to govern and can be smuggled, to computing power itself" — and his identification of the diffusion framework's likely failure mode as assuming a rule adequate today stays adequate as chip performance improves.
    • Six bare-slug wikilinks qualified to their real folders ([[rand-ai-power-requirements]] → [[sources/rand-ai-power-requirements]], and similarly for compute-governance, epoch-ai, export-controls-ai, ai-environmental-impact, americas-ai-action-plan), and two typed relationships added. sources_count 2 → 6; last_updated → 2026-08-04. Confidence held at high, not raised: the rating was unsupported at two sources and is now supported at six, which is a repair of the evidence rather than an upgrade of the claim. Original backed up to _meta/_revision-backups/entities/.

Two claims on the page are marked as resting on search-returned excerpts rather than full reads: the ChinaTalk transcript quotations, and the Journal of Cyber Policy article's characterization of his position on the diffusion rule.

Three house-style violations removed in the process, all in one paragraph: an operational reference to a _meta/ briefing from mainspace (the seven-times-flagged [[_meta/briefings/weekly-2026-W19]] link); the sentence "this entity page was created during the May 10, 2026 continuation pass associated with the v4.0 schema refactor," which is wiki self-reference barred outright by house style §2d; and a pointer to INGEST-nathan-lambert-notes-china-ai-labs.md, a queue task the 08-03 run confirmed was processed and archived. The frontmatter description also carried three wikilinks and a self-referential clause about what the page was "flagged by"; rewritten as a plain description. This removes one of the three mainspace references to weekly-2026-W19; the other two, on AI Macro-Prudential Policy and Index, remain a curator schema question.

Frontmatter repaired (live)

  • 55 of 120 slice-5 entities/ pages had understated sources_count, of which 46 corrections stand after the counting-rule defect described below was found and repaired — 38% of the slice, against 47% on slice 4 and 54% on slice 3. Recomputed from distinct (Source: URL) cites plus distinct wikilinks resolving into sources/. The two largest sit on high-reliance pages: METR, 87 inbound references from 49 pages, declared 10 while citing 21; and National Institute of Standards and Technology (NIST), declared 8 while citing 18. Others of scale: imda-singapore 7→11, nist-caisi 14→17, icrc 1→3, jd-vance 6→8, jake-sullivan 2→4, iapp 1→3. Four pages declared sources_count: 0 while carrying citations: jennifer-king, nsa, herbert-simon, llion-jones.

last_updated was not bumped on the repaired pages, on the precedent set 08-02 and 08-03: a count correction is not a content revision, and bumping the date would reset each page's confidence-decay window and mask real staleness.

The counting rule this run used was defective, and the verification pass caught it. The script counted a wikilink as a citation to a sources/ page whenever the link's basename matched a file in sources/ — including links explicitly qualified to another folder. Forty-eight basenames exist in both sources/ and a content folder (take-it-down-act, colorado-ai-act, texas-traiga and 45 others), so an explicit [[legislation/take-it-down-act]] was miscounted as a source citation. Nine of the 55 corrections were inflated by one each and have been revised back down: henry-farrell 3→2, house-select-committee-on-china 5→4, john-moolenaar 5→4, kalshi 5→4, luiza-jarovsky 8→7, mark-warner 4→3, marsha-blackburn 3→2, naiac 3→2, nairr 3→2. The rule now ignores any link explicitly qualified to a non-sources/ folder. All 57 pages this run touched were re-verified afterwards and agree with the fixed rule exactly.

Two pages declared more than they carry before this run — nita-farahany (6 declared, 5 real) and mustafa-suleyman (7 declared, 6 real). Neither was lowered, per the correct-upward-only rule for pages the run did not otherwise touch. Under a stricter rule that discards every ambiguous bare-slug link rather than resolving it, roughly 50 further slice-5 pages over-declare; the ambiguity is a wikilink-style question (bare slugs where two folders hold the same basename) rather than a count error, and is referred to the curator with the high-confidence audit below.

Queued — foundational sources

  • Guan, Blanchard, Foerster, Jia, Huang, Papernot, "AI Agents Enable Adaptive Computer Worms" (arXiv:2606.03811v1) — score 8, the run's highest. A dangling foundational reference of the purest kind: the 2026-08-03 08:12 developments file carries its quantitative findings and marks the source "not enriched — the preprint itself was not read." Verified: arxiv.org, the canonical host; HTTP 200 with sourceURL, og:url and url all equal to the request, so no redirect to a mirror. Title, six authors, and date agree across the rendered page and four independent metadata fields; the abstract's distinctive passages (the WannaCry comparison, "parasitically uses compromised machines," the zero-marginal-cost claim) appear identically in citation_abstract, og:description and description. arXiv-issued DataCite DOI 10.48550/arXiv.2606.03811. Saved: Raw Sources/AI Agents Enable Adaptive Computer Worms - Guan et al (arXiv 2606.03811, 2026-06-02).md. Queued: INGEST-guan-adaptive-computer-worms-2026.md.

A date divergence the record does not yet carry. The paper was submitted 2 June 2026. It reached the wiki on 3 August 2026 because Jack Clark summarized it in Import AI 467 that day — a two-month lag between publication and surfacing. The developments file's phrasing ("in work summarized on August 3, 2026") is correct and does not commit the send-date-as-event-date error, but nothing in the record states the actual publication date, and an ingest working from the digest alone would have no way to find it. Flagged in the raw file, the ingest task and the verification record.

Three further cautions carried into the ingest task: the reported rates (≈80% detection, ≈53% exploitation, 88% self-replication, ≈37% overall) are not in the abstract and come from Clark's summary; the open-weight model used is not named in the abstract; and the four institutions attributed in the digest (Toronto, Vector Institute, Cambridge, ServiceNow) are unconfirmed against the primary, since the abstract page carries no affiliations. The body was not retrieved and the PDF URL is carried instead.

  • OpenAI Global Affairs, "Keeping America out in front on AI" (2026-08-03) — score 8. OpenAI's own statement of the position it is taking into the August 2026 negotiations, carried in the record only through the developments file. Verified: the publisher's own newsletter, HTTP 200, og:url and sourceURL equal to the request; article:modified_time 2026-08-03. Distinctive-passage check passed on four items, and every quotation the 08-03 08:12 developments file took from its own independent scrape matches this text verbatim. Outbound links resolve to OpenAI-controlled hosts and to a commerce.senate.gov document. Saved: Raw Sources/Keeping America out in front on AI - OpenAI Global Affairs (2026-08-03).md, policy section verbatim in full. Queued: INGEST-openai-keeping-america-out-in-front-2026.md.

The ingest task carries a modality caution: the post describes the administration's action as expected and is conditional throughout, and must not be merged with the separate, Axios-sourced August 3 White House statement that the June 2 order's framework "was complete by the deadline." It also flags that "reverse federalism" is OpenAI's own coinage and must stay attributed.

Authenticity-verification failures

None. Four documents were verified this run — the arXiv preprint, the OpenAI post, the Klobuchar Senate release and Heim's publication list. All four resolved at HTTP 200 on hosts appropriate to their type with corroborating metadata. Three provenance limits are recorded rather than hidden: the arXiv body was not retrieved (abstract page only); Heim's own site was last modified before the 2026 items now on his RAND profile, so it is complete only through January 2025; and the RAND author biography was read through a search-result body rather than a direct scrape of rand.org.

Determined not to be gaps

  • The June 2, 2026 frontier-AI executive order. Referenced from ten mainspace pages and 77 developments files, and the strongest apparent missing-legislation candidate on the raw list. Alias resolution resolved it to the existing EO — Promoting Advanced AI Innovation and Security (Trump, signed June 2, 2026), which carries the signed text. Not a gap — though the slug still says "draft" for an order signed two months ago, which is a rename question for the curator.
  • "Duty of care" as a concept page. Named on 21 mainspace pages and the organizing frame of the August 3 Senate item. AI Liability and AI and Tort Liability both exist and are the landing pages; the new material is a fold, not a page.
  • Palantir and Hugging Face. Both named heavily in the 48-hour window (41 and 7 mainspace pages); Palantir Technologies and Hugging Face both exist.
  • Recent-window items behind the nightly fold, not gaps: Qwen3.8-Max general availability and the weights promise, Palantir's Q2 results, Amazon's $3 trillion market capitalization and completed $35 billion OpenAI tranche, the New York SAFE for Kids implementing rules, the CPPA gig-platform audit, and Lilian Weng's return to OpenAI. Pages exist for each and the developments-log cycle owns them.

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) — the actual policy document behind the CAISI-central-role position, linked from the post queued today and never retrieved. The stronger long-term citation of the two. Score ~6, and the first item for the next run.
  • entities/john-thune — Senate Majority Leader and the Republican half of the vehicle described on August 3, with only two inbound mainspace references. A genuine missing page whose in-degree does not yet justify the cap slot; it will rise as the bill develops. Score ~5.
  • NIST AI 300-1 zero draft (documentation and disclosure guidance, released 2026-07-30) — a standards/ gap, but the only account in the record is a paywalled Inside AI Policy item read as lede only, with the comment deadline not retrievable. Declined on the quality gate pending a direct pull from nist.gov. Score ~5.
  • Gary Marcus, "OpenAI's amazing — but vastly oversold — new model Astra" (2026-08-02) — carried from 08-03, still unqueued, still the strongest deferred essay. Score ~6.
  • New person pages implied by the last 48 hours, none created: levent-alpoge (now load-bearing on three pages after the August 3 half-of-ten reproduction), clement-delangue (7 inbound, Hugging Face CEO, quoted at length August 3 on Chinese open-model dominance), andrew-garbarino (4 inbound, House Homeland Security chairman on the DoorDash letter), jasjeet-sekhon, alex-karp, sabrina-ross, charity-clark, terence-tao, plus the ten carried from 08-03 (devin-kim, isaac-harris, jeremy-pelter, matt-stoller, ernie-davis, henry-yuen, thomas-bloom, noam-brown, lijie-chen).
  • Slice-5 thin anchors not reached, in reliance order after today's recount: jensen-huang (14 inbound, 599 words, confidence: high on 2 real sources), ico (13 inbound, high on 2), matthew-prince (14 inbound, high on 2), knight-first-amendment-institute (16 inbound, 286 words), institute-for-law-and-ai (12 inbound), jam-kraprayoon (10 inbound, 229 words). Score 4–5 each.
  • The high-confidence-on-two-sources pattern is now measurable and is a folder-wide defect, not a page defect. In slice 5 alone, twelve pages carry confidence: high with two or fewer real citations after today's recount — lennart-heim (repaired), jensen-huang, matthew-prince, ico, mariano-florentino-cuellar, lila-shroff, john-jumper, max-tegmark, jony-ive, kevin-mandia, jared-polis, katrina-manson. CLAUDE.md sets a three-source floor for high. This is the fourth consecutive run to raise the question and the first to size it; it is better handled as a single audited pass than as daily gap-scan residue.
  • The slice-5 2026-06-06 cohort. 41 entities/ H–N pages combine in-degree ≥6 with last_updated: 2026-06-06, the date of the v4.3 content/operational split — led by nita-farahany (195 inbound references from 79 pages, the most-relied-upon page in the slice and untouched for two months), luiza-jarovsky (32), jack-clark (24), lennart-heim (22, repaired today), henry-farrell (18). Fourth consecutive slice to report this pattern (24 on slice 1, 8 on slice 2, 35 on slice 4, 41 here — 108 pages across four slices). Plainly folder-wide migration residue.
  • Carried unchanged, genuine but below the reference-count threshold: entities/samuel-weinbach and entities/ilhan-scheer (Aleph Alpha, 2 inbound each); entities/clayton-christensen and entities/adam-smith (2 each, both pre-AI thinkers cited for a framework, still awaiting the curator ruling on whether the wiki wants biography pages for them); the long concept list carried since 07-31; 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 mainspace — down from three pages to two (AI Macro-Prudential Policy and Index) after today's Lambert rewrite. Eighth flag; still a schema question for the curator, not a mechanical rewrite.
  • The bin/lint-scan.py escaped-pipe target-capture bug — eighth consecutive flag. Quantified again today at 82 of 357 targets, 23% of the reported count, putting five nonexistent targets at the top of the list. The one-line fix remains a .rstrip('\\') on the captured target. Outside gap-identifier's remit.

Carried unchanged from prior runs

  • The ingest queue. Two tasks filed today join the NCSL state-AI-tracker task filed by the nightly chain earlier on 2026-08-04; the 08-03 OpenAI mathematics task has been processed and archived. Three open items.
  • 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, Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment), Executive Order 14365 — Ensuring a National Policy Framework for AI, Clawed, California SB 243 — Companion Chatbots, NIST AI Risk Management Framework (AI RMF 1.0).
  • companies/ has never been audited for the sources_count defect. It is rotation slice 0, next due 2026-08-13. On the rates now measured across four slices — 46%, 47%, 54% — and on the fifteen-fold OpenAI understatement found on 08-03, it should be expected to carry the problem at scale.
  • 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; 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 the entity_type-for-for-profit-non-developer schema gap (08-02), which slice 5 reinforces — nextera carries the same inline # comment inside its entity_type field as the five slice-4 pages.
  • Model-slug inconsistency in models/: claude-opus-45, claude-opus-46, claude-opus-47 beside claude-opus-4-8. Fifth flag.
  • The 14 dated dashboard-rebuild-failed notes from June, unrevisited for nine weeks; rebuilds have succeeded on every run since 07-31, including today's. They can be cleared on a curator's word.

Post-run verification pass

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

  • The sources_count counting rule was wrong, and it propagated. Described in full above. It produced a false correction on nine slice-5 pages and, on Amy Klobuchar, a "fix" in the wrong direction: the page was written declaring 4, which was correct, and an earlier version of this report recorded it being "corrected" to 5 on the strength of a miscounted [[legislation/take-it-down-act]] link. Under the fixed rule, and after the citation additions below, the page carries 8.
  • Four claims on the new page were uncited. The TAKE IT DOWN conviction, the AI Whistleblower Protection Act co-sponsor list, the SEARCH Act provisions, and the algorithmic-pricing proposal each rested on a [[legislation/…]] link, which CLAUDE.md does not treat as a citation. Three now carry their own (Source: URL) — the FTC warning-letter item, congress.gov S. 1792, and Schmitt's Senate release — and the conviction now cites both the legislation page and TAKE IT DOWN Act — Source Summary.
  • Two uncited comparative superlatives were removed from Amy Klobuchar: "longest-running Democratic sponsor of federal AI legislation" and "the most frequently recurring Democratic name on bipartisan federal AI bills." Both were contestable rankings with nothing behind them, and both are now plain descriptions. A pairing error was corrected in the same edit — the Cruz pairing is on synthetic intimate imagery, not risk-based accountability, which is the Thune track.
  • A wiki self-reference was introduced on the page whose headline finding was removing three of them. Nathan Lambert gained "cited across the record," where "the record" is the wiki's own corpus. Rewritten to "in the same series."
  • A ## Sources in Wiki section survived the Lennart Heim rewrite. House style §2d bars inbound-link metadata as article content. Removed; its one substantive link, to America's AI Action Plan, was moved into ## Relationships rather than dropped.
  • A provenance limit was recorded on the Washington Post citation, which is a newsletter landing page rather than a dated permalink and was read in the newsletter body; and the stalled status of the Cotton–Klobuchar bill, present on AI Biosecurity and omitted from the new page, was restored.
  • One broken wikilink in an ingest task — [[concepts/offensive-cyber-ai]], which does not exist — replaced with AI and Cybersecurity.
  • Confirmed clean: all wikilinks on the three entity pages resolve (zero broken targets), the wiki-wide normalized broken-target count is unchanged at 269 so the run introduced none, no sourced fact or citation was dropped in either expansion, every one of the six qualified bare-slug targets exists at its qualified path, and confidence was not inflated anywhere.
  • Two report metrics were stated loosely and are corrected here. The word counts compared body text before against whole-file after; body-to-body the expansions are 473 → 1,409 and 369 → 719. And the escaped-pipe defect accounts for 82 of 357 targets (23%), not 88 (25%) — the remaining six are documentation placeholders.
  • Dashboard rebuilt — 1,827 article pages written, no errors.

Findings not acted on, referred onward: neither Index nor _meta/retrieval-index.md was updated for Amy Klobuchar, the systemic gap flagged on 08-03 and outside gap-identifier's remit; and the person, senator, congress, substack, ai-research and ai2 tags in use on these pages are not in any CLAUDE.md tag family, but person alone is entrenched on 105 entity pages, which makes it a schema question rather than a page defect.

One-line summary

One entity page created for the Democratic half of the Senate's frontier-AI vehicle, anchored on a 2023 primary document the record did not carry; the slice's strongest thin anchor tripled in length and its confidence: high finally given the sourcing it claimed; a second page's four house-style violations removed along with a dead queue pointer; 55 slice-5 sources_count errors corrected, including METR at 87 inbound references reporting 21 citations as 10; and two verified foundational sources queued — a June 2 preprint whose findings the record has been carrying second-hand for a day, and OpenAI's own August 3 policy statement.

Categories: Briefings | briefing | gap-scan