Gap Scan — 2026-08-23

Scanned

Recent window: 4 New Developments Log/ files in the last 48 hours (2026-08-21 08:10 and 22:05, 2026-08-22 08:10 and 22:05), 17 items between them. Rotation slice: 10 — analysis/ (all), 26 pages.

Mechanical inputs: bin/lint-scan.py (1,877 content files; 235 distinct broken targets across 259 refs, 24 referenced 2+ times); bin/primary-doc-scan.py --days 3 (6 digests, 3 cited primary-document URLs, 0 missing from Raw Sources/, 3 keyword pre-filter hits read and cleared). 31 candidates generated, 7 actioned.

Gaps actioned (7 of 31 found)

New pages created (live)

• ARC-AGI — score 7 (live thread +3, in-degree ≥10 +2, frontier-capabilities core +2). Gap type 4: the wiki holds the ARC-AGI-2 paper as an ingested foundational source with 13 inbound refs, and names ARC-AGI on 24 content pages and ARC-AGI-3 on 4, but had no anchor page for the series itself — so benchmark scores lived on model pages with nothing defining what they measure. The page covers the skill-acquisition-efficiency definition from Chollet's 2019 "On the Measure of Intelligence," the core-knowledge-priors design constraint, the three generations, the ARC Prize competitions, and the harness-versus-model distinction that the July 2026 Schema result and the August 21 NVIDIA AVO result both turn on. confidence: medium, 5 sources. Depends on queued source: ARC-AGI-3 technical report (provenance note in place).
• Autoriteit Persoonsgegevens — score 5 (live thread +3, AI-regulation core +2). Gap type 3: the Dutch data protection authority was named on exactly 1 content page while "Article 22" appears on 17 and "automated decision" on 44, and it is the GDPR lead supervisory authority for every controller with a Dutch EU main establishment. Covers the one-stop-shop role, its algorithm and generative-AI supervision, and all four Uber fines including the €824,990,000 automated-decision penalty announced August 21, 2026. confidence: medium, 4 sources.

Pages expanded (live)

• Automated Decision-Making Technology (ADMT) — score 5. 438 words, last_updated: 2026-06-18, covering only the California CCPA side of automated decision-making while the largest enforcement action on the concept anywhere had just landed under a different regime. Added an ### The EU comparison subsection setting the CCPA's notice/opt-out/access structure against Article 22's prohibition-plus-exceptions structure, the Uber decision as the worked example, and the 4%-of-turnover ceiling that has no CCPA counterpart. 438 → 625 words, 1 new cite, 2 new typed relationships, eu tag added.
• US AI Regulatory Approaches Compared — score 4 (in-degree 36 → +2, US-regulation core +2). The slice's strongest anchor and its stalest: last_updated: 2026-06-06, 78 days, on a page 36 other pages depend on. Added ### Where the frontier-transparency trigger sits under "What triggers enforcement", covering OpenAI's August 21 request that SB 53 be amended to require monitoring of frontier models under training or evaluation — which would move approach 1's trigger from pre-deployment to mid-training, and puts that approach under pressure from a regulated developer at the same time as EO 14365 seeks to prevent it operating at all. 1,708 → 1,930 words, 3 new cites across two independent sources quoting the post directly.
• Surveillance Pricing — score 3 (live thread +3). Added a ## Municipal legislation section for the Seattle Fair Pricing and Transparency Ordinance — scope, the loyalty-program and volume-discount carveouts, the $3,000/$10,000 penalty tiers, the 25-location private right of action and the $1 million collective-action cap. The same reporting states that Maryland, Connecticut and New Jersey passed state surveillance-pricing regulations in spring 2026, which is broader than the two statutes the page identifies and lists New Jersey as enacted where the page has it as a pending bill; the discrepancy is recorded with attribution and added to ## Open questions rather than resolved in favour of either count. 1,613 → 1,832 words, 1 new cite, 2 new open questions.

Graph wiring

• General Data Protection Regulation (GDPR) — added Autoriteit Persoonsgegevens to the regulated-by enforcement-body list.
• François Chollet — added a depends-on link to ARC-AGI.
• AI Benchmarks and Evaluation — the ARC-AGI-2 bullet now points to ARC-AGI for the series, its definition and the verification process.

Link aliases fixed

None. Ninth consecutive run at zero, and for the reason recorded since 2026-08-18: the referenced-2+ list is author surnames implied by sources/ pages, institution shorthand (entities/carnegie-mellon, entities/uw, entities/ecb), and deliberate forward references. Reported without re-auditing the list, per the skill.

Queued — foundational sources

• ARC Prize Foundation, "ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence" — score 8 (live thread +3, dangling foundational reference +3, core +2). Named on 4 content pages with scores attached; sources/arc-agi-2 exists for the previous generation and nothing for this one. Verified: arXiv:2603.24621, DOI 10.48550/arXiv.2603.24621, corroborated against the ARC Prize Foundation's own PDF and its March 25 launch post. Not downloaded — 795 KB PDF, ~75,000 characters of HTML, above the size where the QUEUE rule prefers a fresh fetch at ingest. Queued: INGEST-arc-agi-3-technical-report.md; verification record in proposed-sources/arc-agi-3-technical-report-2026.md.
• Pew Research Center, "How Much of the Internet Is Written With AI?" — score 8 (live thread +3, dangling foundational reference +3, core +2). Carried in the 2026-08-22 morning digest by its figures alone; "Common Crawl" is named on 18 content pages and "Pew Research" on 9, with no sources/ page for the study. Verified: pewresearch.org, five named authors with staff pages, publication metadata matching the stated date, methodology companion published separately. Saved: Raw Sources/How Much of the Internet Is Written With AI (Pew Research Center 2026).md with both data tables. Queued: INGEST-pew-how-much-internet-written-with-ai.md.

Authenticity-verification failures

• Autoriteit Persoonsgegevens fine decision of 17 August 2026 (Uber) — sought as the primary instrument behind three wiki claims. The AP published a news item on August 21 and no decision document: no PDF, no /documenten entry, no case number, in either the English or Dutch site. The decision's date is attested only through Reuters, which viewed the document; HLN reports that attribution. Rejected at protocol step 5 — full text not on any trustworthy host, and what is reachable is a summary rather than the instrument. Nothing saved, nothing queued. Recorded in needs-review/2026-08-23-ap-uber-decision-not-published.md, which also raises whether a supervisory authority's statement of its own findings should keep the schema's default press-release → supporting classification. bin/primary-doc-scan.py correctly does not flag this: the document has no URL to be missing.

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

• Private Safety Processing — closed, not deferred a fourth time. The 2026-08-22 report named it structural on its third consecutive deferral. It is not a gap: the mechanism is already carried on both OpenAI and Anthropic. What was deferred three times was a page for something the graph had already absorbed. Removed from the backlog rather than re-listed.
• companies/uber — examined and declined, not deferred. "Uber" is named on 31 content pages, which reads mechanically as a large missing-company gap. Reading where: the mentions are illustrative, spread across Aggregation Theory, Agentic Economy, AI Bubble Debate and similar, plus Transportation — AI Deployment and General Data Protection Regulation (GDPR). Uber is not an AI developer and the wiki does not track it as a subject; the August fine belongs where it now sits, on General Data Protection Regulation (GDPR) (folded by the nightly run on 2026-08-22), Automated Decision-Making Technology (ADMT) and Autoriteit Persoonsgegevens. A company page would collect scattered analogies rather than fill a gap.
• Slice-10 stale analysis pages not reached (score 3 each). analysis/interpretability-and-safety (in-degree 6, medium, last_updated: 2026-06-06); analysis/eu-vs-us-ai-regulation (in-degree 9, medium, 2026-07-13); analysis/open-questions and analysis/missing-coverage-2026-04-14 (both sources_count: 0, 2026-06-06, in-degree 6 and 10). The last two are inventory pages whose staleness may be structural rather than fixable; neither carries unchecked planned-action boxes, so the crystallize/lint backlog is genuinely empty this run.
• Live-thread items named in the window with no page (score 3 each). Air Space Intelligence (the Boston startup that won the $875 million FAA SMART contract over Palantir, four Palantir alumni on its executive team; zero content mentions); Amir Salek (Google TPU founder joining Anthropic's compute team, 1 mention); the International Catholic Legislators Network (venue for Pope Leo XIV's August 21 algorithmic-domination address, 0 mentions); Apple Music's "Made With AI" labelling and its March 2026 AI Transparency Tagging predecessor, which folds into AI Content Provenance rather than needing a page.
• Carried from 2026-08-21 and 2026-08-22, unchanged. Vercel (13 mentions, no company page); Marvell (4); models/deepseek-v3 (in-degree 49 on 608 words, routed to model-robustness-check rather than a gap-scan expansion); the remaining slice-9 thin model anchors.
• The 24 referenced-2+ broken targets, unchanged — ninth consecutive run. Twelve are author entity pages implied by sources/ pages at in-degree 2. Two (concepts/inverse-cooking-problem, concepts/inverse-trust-problem) are parked under needs-review as coined-term questions; concepts/algorithmic-decisionmaking is parked as an ADMT-overlap question and was correctly not created here despite this run touching Automated Decision-Making Technology (ADMT) — the expansion made the ADMT page the stronger home, which is option 1 in that note, but the choice remains the curator's.
• The 29 date-stamped headers in lint-report.md (score 1). Unchanged for seven cycles; the report still does not distinguish legitimate procedural histories from timeline dumps.
• The White House release of 2026-07 (score 5–6), instrument still unidentified, carried from the 2026-08-20 primary-document baseline.

Checks run before closing

• No broken outbound links introduced. All wikilinks across the two new pages and six edited pages resolved against the full page set after writing; zero unresolved. One was caught and fixed mid-run: the provenance note on ARC-AGI used [[sources/…]] as illustrative text, which the linter reads as a link; rewritten as prose.
• Broken-link count unchanged at 235 distinct / 259 refs. Expected, and the same pattern as the last four runs: none of the actioned gaps was on the broken-link list, because all were dangling prose references and thin anchors rather than broken targets.
• Style scan clean. All eight touched pages checked programmatically against the banned-header, hyperbole, self-reference, standing-Predictions and date-stamped-header lists. Zero hits on every check; wiki-wide totals unchanged (banned 0, self-refs 0, predictions 0, header hype 0, date headers 29 — none added).
• Revision fidelity verified mechanically, not by eye. For each of the six revised pages, the pre-existing citation URLs and wikilink targets were enumerated from the pre-edit reads and checked against the post-edit files: zero dropped on all six. Every edit was additive; no page was rewritten, so no _revision-backups/ copy was required.
• Cap respected. Seven gaps actioned against a 6–10 cap; 31 candidates found. Eight web pulls and three searches, no crawls.
• Confidence set honestly. Both new pages are confidence: medium; neither met the three-independent-source bar, and ARC-AGI additionally depends on a source not yet ingested. The three expanded pages held at medium rather than rising — Automated Decision-Making Technology (ADMT) gained a fifth source but its subject sits in the six-month standard-decay window on a page whose California half was last reinforced in June.
• Dashboard rebuilt. bin/build-dashboard.py wrote 1,987 article pages.

One-line summary

Two pages created (ARC-AGI, Autoriteit Persoonsgegevens) and three expanded, with two foundational sources queued for the curator's ingest; the Dutch DPA's Uber decision was sought and rejected as unpublished, and two long-running backlog items — Private Safety Processing and companies/uber — were closed rather than deferred again.