Gap Scan — 2026-07-27

Scanned

Recent window: 3 New Developments Log/ files inside 48h (2026-07-25-2204, 2026-07-26-0805, 2026-07-26-2205); 313 wiki pages carrying last_updated in the 07-26 to 07-27 range, dominated by the 07-26 remediation and retrospective-ingest cluster recorded in five consecutive log entries. Rotation slice: 11 — sources/ A–H (306 pages).

Candidate list built from a wiki-wide broken-link scan (run with a corrected target-capture regex that strips the trailing backslash artifact reported by the 07-26 run, which removes the phantom companies/anthropic\-class entries from the ranked list), an in-degree × thinness × frontmatter audit of the full slice-11 set, a dangling-reference scan over the three recent dev-logs, and the standing backlog in queue/gap-scan/GAP-unresolved-link-targets-2026-07-26.md. Deduplicated against the open queue, the gap-scan-* reports from 07-20 to 07-26, and existing Raw Sources/ files.

The slice-11 finding is a reliance/depth mismatch, not a missing-page problem. sources/ A–H holds the four highest-in-degree pages in the wiki — eu-ai-act (316), americas-ai-action-plan (270), california-sb-53 (234), colorado-ai-act (178) — and every one of them was last touched 2026-06-06, sat between 690 and 1,025 words, and was missing the schema-required sources_count field. Twenty-four A–H pages lacked that field outright. The dedicated source-robustness-check skill is the right instrument for the deep cases; this run took the two largest and left the rest as a named backlog for it.

Firecrawl remained unavailable, as on every run since 07-24; the workspace web-fetch and web-search tools were used under the full authenticity protocol (1 fetch and 1 search across 6 gaps, well within cap).

Gaps actioned (6 of ~19 found)

New pages created (live)

None. No genuine missing-page gap cleared the threshold this run. Slice 11 is a sources/ slice, and the autonomy rules bar AUTO from creating sources/ summary pages — those are an output of ingest. The recent-window pageless candidates resolved to existing pages under alias resolution: the CAC Interim Measures for AI Anthropomorphic Interactive Services (07-26 evening dev-log) already has China — Interim Measures for the Administration of AI Anthropomorphic Interaction Services; the Nvidia–OpenAI $250B financing talks, the Samsung–Broadcom $200B MOU and the DeepSeek funding suspension are Snapshot-row updates to existing company pages rather than new pages, and belong to the nightly fold. Per the quality gate, no thin or speculative page was forced.

Pages expanded (live)

• EU AI Act (Regulation 2024/1689) (score ~7) — the squarest gap in the wiki by reliance, and the run's main piece of work. In-degree 316, the highest of any page; 1,025 words summarizing a 107,319-word primary text; sources_count absent; last_updated 2026-06-06, six days before the Regulation's general application date of 2 August 2026. Deepened to 3,777 words directly from Raw Sources/EU AI Act.md — the primary text, not the web. Added: the Article 3(1) definition of an AI system and the Article 3(2) definition of risk, both verbatim; the seven Article 2(1) addressee categories and the output-used-in-the-Union extraterritorial hook; a table of the seven Article 2 exclusions (military/defence/national security, third-country law-enforcement cooperation, sole-purpose scientific R&D, pre-market research with the real-world-testing carve-back, purely personal non-professional deployers, and the free-and-open-source-licence exclusion at 2(12)) plus the DSA intermediary-liability and data-protection savings; the two Article 6 routes to high-risk classification, the Article 6(3) derogation with its four conditions and the profiling override, the 6(4) self-assessment and registration duty, and the delegated-act machinery at 6(6)–6(8); Article 4 AI literacy; Articles 20 and 21 corrective action and cooperation duties; the Article 27 fundamental-rights impact assessment and who owes it; the Article 73 fifteen-day serious-incident deadline; Article 60 real-world testing; the full Article 50 transparency regime including the artistic-work narrowing and the editorial-responsibility carve-out for AI-generated public-interest text; Article 51's two classification triggers with the 10^25 FLOP figure stated correctly as a rebuttable presumption rather than a definition, and the Article 52 two-week notification and rebuttal procedure; the four Article 55(1) systemic-risk duties; all four Article 53(1) GPAI duties; the Article 53(2) open-source GPAI exemption and its exact limits; the Article 53(4) codes-of-practice and harmonised-standards route; the Article 99 penalty architecture as Member-State-implemented with its "effective, proportionate and dissuasive" and SME-viability constraints; and an Article 113 application-date table recording that the Annex I product-safety route at Article 6(1) does not apply until 2 August 2027. Frontmatter repaired (sources_count 2, tags extended). Confidence stays medium — the primary text is slow-decay and now fully read, but the implementation section rests on a single further source, below the three-source bar. Every prior fact, table, tension and citation preserved; the original is backed up in _meta/_revision-backups/2026-07-27/.
• Open-Weight Frontier Models (score ~8) — a high-confidence page carrying a stale fact on the week's dominant thread. The page recorded the "Open Weights and American AI Leadership" letter as having 25 signatories with "OpenAI and Anthropic did not sign." That was accurate on July 24 and is no longer: OpenAI signed on the evening of July 24, Sundar Pichai signed for Google on July 25, and the letter's own page listed 77 signatories when checked today. Corrected without removing the original sentence, which remains true as of its date: added the roster's growth (25 → 32 → 35 → 50 → 77) with the standing instruction to cite the count with its observation date, the 25 named additions, and the confirmation that Anthropic and Amazon are absent at every observation. Added the letter's four substantive positions from the primary text — the right-model-for-the-right-job cost argument, "defenders need access to models with comparable capabilities," the single-points-of-failure argument against capability concentration, and the distillation paragraph, which is the direct industry counter to the Kratsios and Anthropic misappropriation framing already recorded lower on the page. Also added the EU AI Act (Regulation 2024/1689) carve-out material now available: Article 2(12), the Article 53(2) exemption from two of four duties, and its removal at the systemic-risk tier. Provenance note added pointing at the queued ingest. sources_count 24 → 26; confidence stays high. [2 new (Source: URL) cites]
• America's AI Action Plan (score ~6) — in-degree 270, second-highest in the wiki, sources_count absent, stale since 06-06. Unlike the EU AI Act page this one is proportionate to its raw file (852 words against a 2,899-word source), so it was extended rather than rebuilt. Added the plan's own stated rationale for open-weight models under "Encourage Open-Source and Open-Weight AI" — the three reasons given (startup independence from closed providers, sensitive-data adoption by business and government, academic research necessity) and the NTIA-led Commerce convening to drive SMB adoption — which is the standing federal position against which this week's reported administration lean toward case-by-case review of Chinese open-weight models has to be read. Added a typed supports: relationship to Open-Weight Frontier Models and a related: link to the America's AI Action Plan tracker page, which the source page did not previously reference. Frontmatter repaired (sources_count 1, open-source tag). Confidence stays medium.

Corrections (live)

• Anthropic's Responsible Scaling Policy (Version 3.1) (score ~7) — a plain factual error on a heavily-relied page. Its description read "Latest Anthropic RSP." It is not: v3.4 took effect July 8, 2026 and was ingested on 07-25, and v3.2 (April 29) and v3.3 (May 26) sit between them. The v3.4 page already declares supersedes: in the other direction, but 190 files reference v3.1 against 11 referencing v3.4, so the wrong page carries the traffic. Corrected the description; added a "Position in the version sequence" section recording the three later point releases, the fact that v3.4 is the ninth published version, and the four specific elements v3.4 changes rather than inherits — the shift of industry-wide recommendations from ASL-based to argument-based, the Appendix A competitor-contingent commitments that v3.1 does not contain, the Frontier Safety Roadmap's explicit non-commitment status, and the Risk Report cadence, off-cycle triggers, redaction grounds and external-review commitment. Added the reciprocal related: link to v3.4 and links to v2.2 and Responsible Scaling Policy (RSP). Frontmatter repaired (sources_count 2, tags extended). Left for the curator: whether this page should formally take status: superseded with superseded_by / superseded_date is a supersede-skill decision, not gap-identifier's; it was not taken here. The page documents a real April 2026 document faithfully, so the argument for leaving it active is not weak.

Frontmatter repaired (live)

• sources_count added to 24 slice-11 sources/ pages that lacked the schema-required field: ab-2013-training-data, adolescence-of-technology, ai-lead-act, anthropic-rsp-v22, broad-timelines, california-ccpa-regulations, california-sb-243, california-sb-53, chatgpt-content-moderation, china-us-different-ai-races, chinchilla-paper, claude-mythos-system-card, claude-sonnet-46-system-card, clawed, emergence-ai-ethics-auditing, emotion-concepts-llm, eo-13859, eo-14365, frontier-compliance-framework, generative-ai-at-work, geopolitics-in-the-age-of-ai, gpt-53-codex-system-card, gpt-54-thinking-system-card, guide-to-ai-in-agentic-era. Counts derived as one (the anchoring raw source) plus the number of distinct inline (Source: URL) citations on the page — 1 for eighteen of them, 2 for six. last_updated deliberately not bumped: no substantive content changed.

Link aliases fixed

None. Re-ran alias resolution against all folders on every broken target with ≥2 inbound references. The list is unchanged in character from the 07-20 → 07-26 runs and contains no genuine alias errors: literal illustrative text inside provenance notes (sources/... at 8, wikilinks at 4 — intentional prose, not links); _meta cross-namespace links from legacy pages (briefings/weekly-2026-W19 at 3); standing curator-blocked carries (claude-code at 7, claude-cowork at 2); Raw Sources/… path references used deliberately as provenance; and person/institution pages at in-degree 2 that are genuine missing-page candidates below the action threshold. Applying the corrected regex removed the pipe-escape phantoms the 07-26 run diagnosed, confirming that diagnosis — the underlying bin/lint-scan.py fix remains lint's lane and is still unapplied.

Queued — foundational sources

• "Open Weights and American AI Leadership" (multi-company industry letter, July 24, 2026) (score ~8: live thread +3, dangling foundational +3, wiki core area +2) — gap type 1. Named in prose on Overview and Open-Weight Frontier Models with supporting news cites only; deferred as a source-queue candidate on 07-25 pending the fold, which has since happened. The open-weight restriction fight is the single most active thread in the last 72 hours of dev-logs. Verified: microsoft.com/en-us/corporate-responsibility/topics/open-weight/ — canonical header and og:url match the request URL, og:title matches, article:modified_time 2026-07-26T22:31:20Z, body dateline "July 24, 2026", and four distinctive passages present verbatim ("institutional sovereignty"; "openness may be one of the most important paths to AI safety and security"; "keeping the frontier plural"; the distillation paragraph). Corroborated by Tom's Hardware (07-24, 25 signatories), NYT (07-25), Forbes (07-25), and the wiki's own 07-25 and 07-26 dev-log entries, which independently assert the same launch date and the same Anthropic/Amazon absences. Saved: Raw Sources/Open Weights and American AI Leadership - Industry Letter.md. Queued: INGEST-open-weights-american-ai-leadership-letter-2026-07-27.md. Verification trail: queue/gap-scan/proposed-sources/open-weights-american-ai-leadership-letter-2026.md.
  • Provenance caveat carried into the raw file, the INGEST task and the trail: the letter is hosted by Microsoft, one of its signatories, not by a neutral coalition secretariat. No single entity owns it and no secretariat domain exists. A signatory's own corporate domain is not an aggregator, so the protocol's primary-host preference is satisfied, but the arrangement should be stated on the resulting source page. A second site, openweights.gitlawb.com, was found and not used: unattributed domain soliciting signatures rather than publishing the authoritative roster.
  • Fast-decay warning carried into the INGEST task: the signatory count moved 25 → 32 → 35 → 50 → 77 across four days, and the wiki's own record (32, from the 07-26 dev-log) was already behind. Recommend citing the roster with its observation date.

Authenticity-verification failures

None. The single pull resolved on a signatory's own domain with matching title, date, and four distinctive passages.

Tooling note. The firecrawl family has now been unavailable on four consecutive runs (07-24 through 07-27). The workspace web-fetch tool returned the full rendered letter body and the complete 77-name roster rather than a model-written summary; the saved raw was checked against the returned text before writing and carries fetched: / fetch_note: frontmatter recording the method. Worth the curator's attention as a standing tooling gap rather than a per-run inconvenience.

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

• Slice-11 residue — the strongest remaining backlog in the wiki, and the natural queue for source-robustness-check rather than for gap-identifier. Reliance/depth mismatches, all stale since 2026-06-06 unless noted: California SB 53 — Transparency in Frontier AI Act (in-deg 234, 740w — the largest untouched mismatch left, ~6); Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment) (in-deg 178, 690w, ~5); Executive Order 14365 — Ensuring a National Policy Framework for AI (in-deg 144, 857w, ~5); Clawed (in-deg 105, 1,085w, ~4); California SB 243 — Companion Chatbots (in-deg 98, 415w — the worst words-per-inbound-link ratio in the slice, ~5); China and the US Are Running Different AI Races (in-deg 89, 662w, ~4); Executive Order 14110 — Safe, Secure, and Trustworthy AI (in-deg 82, 728w, ~4); AI LEAD Act (S. 2937) (in-deg 66, 1,092w, ~3); China — Interim Measures for the Management of Generative AI Services (in-deg 61, 591w) and BIS Framework for AI Diffusion — Interim Final Rule (Jan 13, 2025) (in-deg 61, 910w, ~4 each); G7 Hiroshima Code of Conduct for Advanced AI (2023) (in-deg 57, 511w, ~3); Alignment Faking in Large Language Models (in-deg 55, 621w, ~3); Emotion Concepts and their Function in a Large Language Model (in-deg 55, 493w, ~3); China — Provisions on the Administration of Deep Synthesis Internet Information Services (in-deg 52, 599w, ~3); The Adolescence of Technology (in-deg 52, 1,371w, ~3). Six of the fifteen are medium confidence at in-degree above 50.
• Frontmatter question for lint, not for gap-identifier. 175 of the 306 A–H sources/ pages — 57% of the slice — carry confidence: high on sources_count: 1. The decay table sets the high bar at "3+ sources, no contradictions, reinforced within the decay window," which a single-source summary page cannot in general meet. At 57% this is plainly a systematic convention inherited from ingest, not 175 individual errors, and the fix is a policy decision rather than a cleanup task: either the bar is different for a sources/ page that faithfully summarizes one primary document — in which case CLAUDE.md should say so explicitly — or a majority of the slice is overrated. Flagged rather than actioned; a wrong call here propagates to every folder. (Related: source_class is absent from every sources/ page; per CLAUDE.md that field belongs to the raw file, so this is a scanner false positive, not a gap.)
• Recent-window items behind tonight's fold, not gaps yet: the Nvidia–OpenAI ~$250B financing guarantee for the 10 GW southern Ohio data center (Snapshot rows on Nvidia & TSMC — AI Compute Infrastructure and OpenAI); the Samsung–Broadcom $200B MOU through 2030 on 2nm and HBM4E/HBM5; DeepSeek's July 25 suspension of its ~$1.5B round at a $71B valuation and the unverified Liang Wenfeng transcript behind it — the transcript remains the standing authenticity rejection from 07-24 and should stay rejected absent a verifiable host; Google's Fluidstack investment and the Baton Rouge / Colorado City Anthropic backstops; the Zvi Mowshowitz July 26 follow-up and the Tech Policy Press CFR/Stanford discussion of the Hugging Face incident's governance consequences (both essay/source-queue candidates, ~4); the college-admissions-essay homogenization study (a concepts/ candidate at ~3, and the underlying paper is a foundational pull); Musk's Economist interview on Chinese AI leadership.
• Genuine missing-page candidates confirmed by alias resolution but below the in-degree threshold (2 each), carried from prior runs: concepts/foundation-models, concepts/economic-possibilities-for-artificial-intelligence, concepts/algorithmic-decisionmaking, concepts/ai-and-language-models, concepts/talent-flow-china-us, concepts/data-broker-regulation, concepts/a-vision-of-democratic-ai, concepts/safety-training-methodologies, concepts/ai-coarse-grainings, concepts/cross-national-ai-policy-tracking; person pages entities/anil-seth, entities/orin-kerr, entities/eric-goldman, entities/kevin-klyman, entities/lee-anne-fennell, entities/michael-j-d-vermeer, entities/sacha-altay, entities/joseph-bernstein, entities/roge-karma, entities/alondra-nelson, entities/clayton-christensen, entities/adam-smith, entities/dina-powell-mccormick, entities/raja-krishnamoorthi, entities/samuel-weinbach, entities/ilhan-scheer; institutions entities/federal-reserve, entities/ecb, entities/bank-of-england (all three from International Monetary Fund (IMF)), entities/carnegie-mellon, entities/santa-fe-institute, entities/sais, entities/uw.
• bin/lint-scan.py target-capture bug — still open. The 07-26 run diagnosed it and proposed the fix (\[\[([^\]|#]+?)(?:\\?\|[^\]]*)?\]\], or strip a trailing backslash from the captured target). Confirmed correct here by running the corrected regex. Unapplied; lint's lane.
• Carried from prior runs: slice-10 residue (US AI Regulatory Approaches Compared at in-deg 28, Interpretability and Safety: From Microscopes to Arms Races, EU vs. US AI Regulation: A Deep Comparison, and the high-on-sources_count: 0 working documents Missing Coverage: Companies, People, Models, Legislation, Concepts (revised), Recommended New Sources for Raw Sources/ (2026-04-14), Manual Ingest Queue); slice-9 model/industry residue (DeepSeek-V3 in-deg 48 at 608w, DeepSeek-R1, Qwen3, Kimi K2, Gemma (Google open-weight models), Helix (Figure AI Vision-Language-Action model), LongCat-2.0 (Meituan), IsoDDE (Isomorphic Labs), Defense / Military — AI Deployment, Retail — AI Deployment, Energy and Electric Power Sector, Construction — AI Deployment); industries/manufacturing topical hole; slice-8 residue (Character.AI Litigation (umbrella) high on sources_count: 0; DHS — Department of Homeland Security (AI Deployer); Defense Innovation Unit (DIU); misratings on ICE — Immigration and Customs Enforcement (AI Deployer), OECD Recommendation on Artificial Intelligence, DARPA (Defense Advanced Research Projects Agency)); H.R. 9619 primary text (congress.gov empty body ×4); METR "expenditure horizon", Zitron "Subprime Data Center Crisis", Pethokoukis transformative-AI essay; the OpenAI/Apollo "metagaming" post and the four arXiv papers surfaced by the 07-26 Redwood pulls; METR's May 19 2026 frontier risk report; Substack-redirect URL upgrades.
• Standing carries (needs-review): claude-code / claude-cowork placement (curator-blocked, 07-09); inverse-cooking / inverse-trust coined terms (07-10); companies/fairly-trained misfile (07-16); entities/cdao / government/cdao duplicate (lint's lane); the 14 dated dashboard-rebuild-failed notes from June, which no run has since revisited.

One-line summary

Slice 11 turned out to hold the wiki's four most-relied-upon pages, all frozen since June 6 and all missing a required frontmatter field: EU AI Act (Regulation 2024/1689) was rebuilt from the primary text at in-degree 316 (1,025 → 3,777 words, adding the scope exclusions, the two high-risk classification routes, the Article 50 transparency regime, the correct rebuttable-presumption reading of the 10^25 FLOP threshold, and the application-date table), Anthropic's Responsible Scaling Policy (Version 3.1) was corrected off its "Latest Anthropic RSP" claim onto the v3.4 that has been in force since July 8, Open-Weight Frontier Models was corrected off a 25-signatory count that is now 77, America's AI Action Plan gained the plan's own open-weight rationale, 24 pages had sources_count restored, and the industry letter at the centre of the week's open-weight fight was verified on a signatory's own domain and queued for the curator's ingest — with a note that 175 of 306 A–H source pages sitting at high confidence on a single source is a schema question the curator should settle rather than a backlog to grind through.