Gap Scan — 2026-08-06

Scanned

Recent window: 4 New Developments Log/ files (2026-08-04 08:11, 2026-08-04 22:05, 2026-08-05 08:09, 2026-08-05 22:12). Rotation slice: 7 — legislation/ (all), 132 pages. Broken-link analysis across 1,757 mainspace pages: 255 distinct broken targets, 285 instances, after stripping the backslash-escape artifacts that have headed the raw distribution since 07-30.

The window and the slice converged: the Senate Commerce executive session of August 5 moved four AI-relevant bills on the same day the rotation reached the folder those bills belong in. Six of the seven gaps actioned come out of that overlap or out of the standing deferred backlog.

Gaps actioned (7 of 23 found)

New pages created (live)

• Youth AI Privacy Act (S. 4199) — score 7 (live thread +3, wiki core area +2, rotation slice +2). S. 4199, sponsored by Ed Markey, ordered reported by voice vote on August 5. The record carried it in exactly one sentence, inside State-Level AI Regulation, as one of five bill numbers in a markup list. Alias resolution confirmed no page under any variant. Built from congress.gov (sponsor, introduction date, referral), the committee's own release (the vote, the five adopted amendments, Cruz's prepared description), Broadband Breakfast's markup account, and EPIC's endorsement and post-markup posts. The substantive finding is that the bill changed shape in committee: the substitute removed the FTC rulemaking authority in the introduced text and replaced it with a statutory 30-day default retention cap extendable by verifiable parental consent, and — per EPIC — the private right of action came out. confidence: medium, 5 sources.
• Children's Artificial Intelligence Toy Safety Act of 2026 (S. 5171) — score 7. S. 5171, Duckworth and Murkowski, introduced July 29 and ordered reported with a substitute on August 5. Zero mainspace mentions before today. Built primarily from the full introduced bill text on congress.gov, so the page carries the operative provisions rather than a characterization of them: the seven-item National Academies study mandate, the two-year FTC/CPSC joint action plan with its four reporting committees, and the complete Section 2 definitions table. The definitions are the part worth having — "child" is drawn at under 14, narrower than KOSA's under-17, and the AI-chatbot definition turns on open-ended unscripted output with five enumerated carve-outs. Duckworth's office supplies the regulatory-gap framing: no federal safety standard reaches the AI inside a toy, only choking, lead and flammability, and CPSC told Congress earlier in 2026 that it lacks explicit authority over non-physical harm. confidence: medium, 4 sources.

Pages expanded (live)

• EU General-Purpose AI Code of Practice (2025) — the strongest slice-7 thin anchor and a fast-decay failure that had actually come due. 96 inbound references, confidence: high on a declared sources_count: 2, zero inline citations anywhere in the body, untouched since 2026-06-06 — and its central forward-looking sentence read "Commission GPAI enforcement powers, including fines, activate 2 August 2026," a date that passed four days before this scan. Added an ## Enforcement from 2 August 2026 section built on the Commission's own 31 July 2026 press release and its regulatory-framework and enforcement pages: that enforcement began on schedule; the AI Office's specific powers over GPAI models; the Article 88 exclusive-competence and Article 89(1) monitoring basis (the provision that gives the Code its supervisory role); the Articles 91–93 non-fining powers and the Article 101 fining power at up to €15 million or 3% of worldwide turnover; that refusing an Article 91 documentation request or Article 92 model-access request is itself finable; the three new reporting channels (complaints tool, whistleblower tool, downstream-provider channel); and the parallel transparency rules with their separate Code of Practice on AI-generated content, first signatory list of 180+ organisations. The obligations/enforcement distinction is now explicit — Chapter V duties applied from 2 August 2025, enforcement only from 2 August 2026. 993 → 1,310 words, 4 new (Source: URL) cites. Frontmatter repaired: sources_count 2 → 14 on a recount (4 new inline cites plus 10 distinct wikilinks resolving into sources/). The page's high rating was already substantiated by those ten source-page links; what was wrong was the declared count, which understated them fivefold. Confidence held at high, not raised. Fidelity check run against the backup at _meta/_revision-backups/legislation/gpai-code-of-practice.2026-08-06.md: 0 wikilinks dropped, 0 citations dropped, 0 of 18 checked factual anchors dropped.

Link aliases fixed

None required. The one candidate that looked like a missing-page gap resolved: "the CHATBOT Act" (S. 4407), named on 10 mainspace pages, is already CHATBOT Act (Cruz–Schatz–Curtis–Schiff, April 2026) — State-Level AI Regulation had piped the alias correctly, and the committee release confirms S. 4407 is the Cruz–Schatz–Curtis–Schiff bill that page describes. Not a gap. (The page's slug says "cruz-parental-controls-bill" for a bill whose short title is the CHATBOT Act, which is a rename question for the curator, not a gap-scan action.)

Queued — foundational sources

• OpenAI, "Democratic Governance of Frontier AI: A blueprint for a federal framework" (2026-06-02) — score 8. Twice deferred, and named by the 08-05 scan as an item that "should take a cap slot on 08-06 regardless of what the window produces." It is the primary text behind OpenAI's CAISI position, its preemption ask, and the term reverse federalism that Reverse Federalism already tracks — all of which the record has carried through a newsletter post and news restatement. Verified: cdn.openai.com, OpenAI's own CDN; HTTP 200, 9 pages; sourceURL and url equal to the request, no mirror redirect; PDF internal title "Frontier safety blueprint." Corroborated twice independently — Haug Partners LLP cites the identical URL, title and June 2 date, and Kapoor and Narayanan link the same URL from normaltech.ai six weeks later. Distinctive-passage check passed on four items including the "which we call reverse federalism" coinage and "CAISI's role should be to conduct evaluations and recommend mitigations—not to approve or block deployments." Saved: Raw Sources/OpenAI - Democratic Governance of Frontier AI, A Blueprint for a Federal Framework (2026-06-02).md, full text. Queued: INGEST-openai-blueprint-federal-framework-2026.md. Verification record: queue/gap-scan/proposed-sources/openai-blueprint-federal-framework-2026.md.
• Kirgis, Kapoor, Schwartz, Rabanser et al., "Can AI agents conduct open-ended AI research? Early evidence from two case studies" (arXiv:2607.27191) — score 8. A dangling foundational reference from the live window: the 08-05 08:09 developments file carries the method, the five failure modes and the 24-author list in detail with nothing behind them. Verified: arxiv.org, the canonical host for a preprint; HTTP 200; submitted v1 2026-07-29 17:57:19 UTC; DataCite DOI 10.48550/arXiv.2607.27191; the 24-name citation_author list matches the authors' own published list name for name and in order; submitting author is co-author Sayash Kapoor. Distinctive-passage check passed on five items. Saved: none — URL-only queue (1,453 KB PDF; the released expert reviews, survey responses and logs come through better fetched fresh). Queued: INGEST-kirgis-shadow-evaluations-open-ended-ai-research-2026.md. Verification record: queue/gap-scan/proposed-sources/kirgis-shadow-evaluations-2026.md.
• Gary Marcus, "OpenAI's amazing — but vastly oversold — new model Astra" (2026-08-02) — score 6. Deferred four consecutive runs; read in full by the 08-03 scan, held back only for cap, and carried since as "the strongest deferred item." Verified: garymarcus.substack.com, the author's own publication, dated Aug 02 2026 on both the post page and the archive index; independently republished in full by ACM at cacm.acm.org/blogcacm/, a scholarly publisher rather than an aggregator. Four-part structure, subtitle, opening line and the two embedded Aug 1 posts are identical across both hosts. Saved: none — URL-only queue, with the ACM copy carried as a fallback host in case the Substack meters. Queued: INGEST-marcus-astra-vastly-oversold-2026-08-02.md. Verification record: queue/gap-scan/proposed-sources/marcus-astra-vastly-oversold-2026.md.

Authenticity-verification failures

None. All four sources pulled this run verified on their canonical hosts with independent corroboration.

One provenance deviation recorded rather than left implicit: two of the three queued sources were saved URL-only. The arXiv paper fits the established large-PDF pattern. The Marcus essay does not — it is a newsletter post that could have been captured in full, and was not, because of the run's cap. It is well-verified on two hosts and the ingest task carries both; the deviation is noted so it is visible rather than silently absorbed.

A date error already in the record

The 2026-08-05 08:09 developments file dates the shadow-evaluations findings to "August 5, 2026." That is the date of the authors' Substack summary. The paper was submitted to arXiv on 2026-07-29 — a seven-day slip, and precisely the send-date-as-event-date failure class named in source-fidelity and in the v4.7 schema changelog. It is flagged in both the ingest task and the verification record so the sources/ page carries 2026-07-29 for the paper and 2026-08-05 for the essay. Any page that later cites the study should be checked for the same slip.

Two substantive facts also appear in the paper's abstract but not in the dev-log item, and should be picked up at ingest: the two shadowed papers were unpublished NeurIPS 2026 submissions, and a robustness check with a second model and scaffold reproduced the failures. The second materially strengthens the finding.

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

• companies/discovery-loop — score ~5, and the strongest deferred new-page candidate. Jeff Dean left Google on August 5 after 27 years to co-found Discovery Loop, an independent public benefit corporation automating machine-learning, science and engineering research, joined by Google senior fellow Sanjay Ghemawat, Google DeepMind senior research scientist Oriol Vinyals, and Google Brain founding member Quoc Le, with Alphabet as a founding investor supplying cloud compute. Four of the most senior names in the field leaving one company together is a substantial event and Jeff Dean already exists. Held on the quality gate: the company is one day old with zero mainspace mentions and no independently verified sources pulled this run. Its in-degree will rise once the nightly fold lands, and it should clear the cap on 08-07 rather than be built today from a single unverified digest item.
• Koray Kavukcuoglu (1 mention) and the Google DeepMind leadership change — Demis Hassabis moving to chair of Google DeepMind and chief scientist of Alphabet, Kavukcuoglu taking daily operations as SVP reporting to Pichai. A fold for the developments-log cycle onto Google DeepMind and Demis Hassabis rather than a new page, except for Kavukcuoglu himself, who now runs a frontier lab and has no page. Score ~4.
• Meta's Muse Code (0 mentions) — Meta's first coding agent, preview released August 5 at $1.25/$4.25 per million tokens. models/muse-spark exists; whether an agent product gets a models/ page is the same unresolved question as claude-code/claude-cowork, on the needs-review list since 07-09. Not actioned.
• Slice-7 pages combining high reliance with a stale last_updated and no inline citations, in reliance order: bletchley-declaration (100 inbound, 5 source-links, 0 inline cites), seoul-frontier-ai-safety-commitments (98, 9 source-links, 1 inline), eo-14110 (88, 14, 2), take-it-down-act (83, 7, 1), china-generative-ai-interim-measures (60, 5, 0), bis-ai-diffusion-rule (58, 5, 0), paris-ai-action-summit-declaration (56, 6, 0), hiroshima-code-of-conduct (55, 9, 0), sb-1047 (48, 12, 0). All are adequately sourced through wikilinks and none is a confidence defect; each is an update candidate on staleness alone, all nine sitting at last_updated: 2026-06-06. Score 3–4 each.
• legislation/gdpr (24 inbound, 2 real sources) and legislation/nebraska-lb525 (16 inbound, 2) — the two genuine high-on-thin-sourcing cases in the slice with meaningful in-degree, after the recount above. Score ~3 each.
• entities/steven-adler (6 inbound, 126 words, low on 1 source, untouched since 07-16), entities/susie-wiles (7 inbound, 330 words) and entities/sanjog-misra (7 inbound, 270 words) — slice-6 carries, unreached for a second run.
• New person pages implied by the window, none created: clement-delangue and levent-alpoge (both ~7 inbound, both named on 08-05 as matching Garbarino's in-degree at the time he was actioned), peter-kirgis (first author of the paper queued today), ernie-davis (the critique unique to the Marcus essay queued today), plus terence-tao, noam-brown, jasjeet-sekhon, alex-karp, sabrina-ross, charity-clark, and the ten carried since 08-03. tammy-duckworth and lisa-murkowski join the list today — both are named sponsors on a page created this run and neither has an entity page, which is why the toy-act page names them in plain text rather than linking them.
• entities/epic, entities/cpsc, entities/ccia, entities/national-academies — four organizations load-bearing on the two pages created today with no pages of their own. EPIC in particular is now cited three times on Youth AI Privacy Act (S. 4199) and authored the model bill both it and People-First Chatbot Act draw on. Score ~4 for EPIC, ~3 for the others.

Standing carries (repeat findings, not per-page gaps)

• The high-confidence-on-thin-sourcing pattern is markedly better in legislation/ than in the entities/ slices — and a first-pass measurement in this run got that backwards. An initial count using a narrow definition of a "real" source (only [[sources/…]]-prefixed wikilinks) put 51 of 132 slice-7 pages at confidence: high on two or fewer sources. That figure is wrong and is retracted here. Recounted under the rule the prior scans actually use — distinct (Source: URL) cites plus distinct wikilinks of any form resolving into sources/ — the correct figure is 12 of 132 pages (9%), against 12 in slice 5 alone out of a much smaller candidate set. Most legislation/ pages cite by bare-slug wikilink into sources/ rather than by inline URL, which the narrow metric could not see: sb-1047 resolves to 12 source pages, eo-14110 to 14, seoul-frontier-ai-safety-commitments and hiroshima-code-of-conduct to 9 each. Their high ratings are substantiated. The residual 12 — led by gdpr (24 inbound, 2 real) and nebraska-lb525 (16 inbound, 2) — are genuine and small enough to fix individually rather than as a folder pass. The methodological point is the durable one: any future audit of this pattern must count bare-slug wikilinks into sources/, or it will manufacture a folder-wide crisis out of a citation-style convention.
• What legislation/ does have folder-wide is zero inline citations. 11 of those 12 pages carry no (Source: URL) at all, and the pattern extends well beyond them — EU General-Purpose AI Code of Practice (2025) had 96 inbound references and not one inline cite before today. This is a style divergence rather than a sourcing deficiency, but it makes fast-decaying claims unverifiable at a glance, which is exactly how the 2 August enforcement date sat stale on the most-linked page in the slice.
• sources_count understatement, slice 7: 11 pages with in-degree ≥5 declare fewer sources than they cite, including five that omit the field entirely — california-sb-53 (242 inbound, field missing, 5 real), new-york-raise-act (123, missing, 3), eo-trump-ai-cyber-2026-draft (70, missing, 27 real — the largest single understatement measured in any slice), colorado-sb-189 (15, missing, 4), pax-silica (7, missing, 5). Also eu-ai-act 20→25 and eo-14365 5→7. Not corrected in bulk this run; the three pages touched today were corrected individually.
• The 2026-06-06 cohort reaches legislation/: 54 pages combine in-degree ≥6 with last_updated: 2026-06-06, led by gpai-code-of-practice (96, repaired today), take-it-down-act (83), china-generative-ai-interim-measures (60), bis-ai-diffusion-rule (58). Sixth consecutive slice to report this (24, 8, 35, 41, plus slice 6, now 54 — over 160 pages across six slices). Migration residue from the v4.3 content/operational split.
• companies/doordash placement — third run to say the curator ruling is what is missing, not the research. Unchanged.
• The bin/lint-scan.py escaped-pipe target-capture bug — ninth consecutive flag. The one-line fix remains a .rstrip('\\') on the captured target. Outside gap-identifier's remit; this scan works around it locally.
• [[_meta/briefings/weekly-2026-W19]] cited from two mainspace pages — ninth flag; a schema question for the curator.
• Model-slug inconsistency in models/: claude-opus-45, claude-opus-46, claude-opus-47 beside claude-opus-4-8. Sixth flag.
• Needs-review standing carries: claude-code/claude-cowork placement (07-09); inverse-cooking / inverse-trust coined terms (07-10, curator-blocked, not re-actioned); companies/fairly-trained misfile (07-16); the entities/cdao / government/cdao duplicate; models/mai-cyber-1-flash decline (07-28); Andon Labs entity_type (07-30); the Situational Awareness LP placement and the entity_type-for-for-profit-non-developer schema gap (08-02).
• 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).
• The ingest queue was empty at the start of this run — every task from the 08-04 and 08-05 scans has been processed and archived. The three filed today are the only open items.

One-line summary

Two Senate bills that moved on August 5 now have pages built from primary text rather than a bill-number list — Youth AI Privacy Act (S. 4199) and Children's Artificial Intelligence Toy Safety Act of 2026 (S. 5171); the EU GPAI Code page, relied on by 96 pages and carrying zero citations, was repaired and now records that Commission enforcement actually began on 2 August rather than predicting it; and three long-deferred foundational sources are verified and queued for the curator's review, one of them carrying a seven-day date error already sitting in the record.