Gap Scan — 2026-08-21

Scanned

Recent window: 4 New Developments Log/ files (2026-08-19 08:05 and 22:08; 2026-08-20 08:11 and 22:09), 34 items. Rotation slice: 8 — litigation/ + standards/ + government/ (82 pages; 52 litigation, 10 standards, 20 government).

Candidate generation: bin/lint-scan.py broken-wikilink report (235 distinct targets, 259 refs; 24 referenced 2+ times, worked as the candidate list per the skill); bin/primary-doc-scan.py --days 3 (6 digests, 0 cited primary-document URLs missing, 8 named-lane pre-filter hits); the lane-1 backlog in queue/gap-scan/needs-review/2026-08-20-uncaptured-primary-documents.md; in-degree × thinness across the rotation slice; and a recent-thread pass over the four digests.

Thirty-four candidates survived deduplication against the open queue, the last seven days of gap-scan reports, and existing Raw Sources/ files. Eight were actioned.

Gaps actioned (8 of 34 found)

New pages created (live)

• Unitree Robotics — score 5 (live thread +3, national-security/export core +2). Unitree was named in the 2026-08-19 08:05 digest with a STAR Market debut, an FCC Covered List restriction and a Defense Department listing attached to it, and had no page. Alias resolution found no existing page covering it: companies/ holds clone-robotics, figure-ai and sunday-robotics, none of which is Unitree, and FCC Covered List addition — foreign-produced advanced robotic devices and connected power inverters (July 2026) describes the restriction category without naming the company. Built from Reuters, The Robot Report, Yahoo Finance and the Washington Post: the August 19 listing (845 yuan close, 460% above the 150.8 yuan IPO price, roughly $50 billion market value, ~6.1 billion yuan raised on about 10% of shares), three years of profit figures, humanoid unit sales rising from 5 in 2023 to 5,215 in 2025 with about 70% going to universities and research institutions, the G1 at $13,500, the "Superman" preview, the July 2026 FCC restriction on new models, the June 2026 Defense Department listing, and attributed criticism from two competing-firm chief executives, labelled as interested-party comment rather than independent assessment. 1,050 words, confidence: medium, 5 sources.

• News Bargaining Incentive (Australia, 2026) — score 3 (live thread +3). Passed August 20, 2026 and carried in that evening's digest. "News Media Bargaining Code" returned zero hits anywhere in the content folders, so the wiki held neither the 2021 Code nor its successor; legislation/australia-voluntary-ai-safety is the only Australian page and is unrelated. Built from the Australian Treasury ministers' joint media release (canonical) plus The Next Web and Reuters: the August 13 introduction and August 20 passage, the A$250 million and "significant service" scope test, the four covered companies, the rate's movement from 2.25% to 2.5% with LinkedIn's addition in early August, the eight-publisher offset minimum, the 150%/200% offset weighting and the 25% single-deal cap, the reporting-period timing rule, the News Journalism Payments scheme and small-publisher grant layer, the relationship to the 2021 Code that Meta exited in 2024, and Meta's unwithdrawn free-trade-agreement objection. Records that AI companies are outside the scheme as passed. 900 words, confidence: medium, 4 sources.

Pages expanded (live)

• DHS — Department of Homeland Security (AI Deployer) — score 4 (in-degree 14 → +2; US AI policy core +2). The highest in-degree page in the rotation slice and the thinnest relative to that: 438 words, sources_count: 4, last_updated: 2026-06-06, and carrying confidence: high on a body that predated the January 2026 inventory entirely. Expanded to 1,462 words with 8 new supporting sources: a new ## The AI Use Case Inventory section (the statutory and executive basis DHS cites — the Advancing American AI Act, EO 13960, OMB M-25-21; the January 28, 2026 release recording 200+ use cases, close to 40% above July 2025; ICE driving the growth; the Brennan Center's finding that the inventory is partial; the American Immigration Council's competing reading of the mid-2025 update as reorganisation rather than reduction); ICE use cases DHS-2457 and DHS-2458 with DHS's own stated safeguards and its own recorded potential impact; Palantir's first appearance in the inventory; the Clearview AI contracts at HSI and CBP; the Flock Safety licence-plate route; CBP use cases DHS-311, DHS-2565 and DHS-2570; Mobile Fortify and the documented double misidentification; the 23 biometric applications; the withdrawn facial-recognition policy; and a new ## Oversight and controversy section. Fidelity checked mechanically against the pre-revision backup: zero citations dropped, zero wikilinks dropped, zero of 20 probed original facts missing. Backup at Wiki/_meta/_revision-backups/government/dhs.2026-08-21.pre-gapscan.md. confidence left at high — reinforced within its decay window, not inflated; sources_count 4 → 12.

Link aliases fixed

None. Fifth consecutive scan reporting zero, consistent with the skill's standing finding: the referenced-2+ list is author surnames, institution shorthand and deliberate forward references, not aliases for existing pages. Not re-audited beyond the alias-resolution pass on this run's own candidates.

Queued — foundational sources

Five, of which four are the lane-1 backlog that needs-review/2026-08-20-uncaptured-primary-documents.md explicitly assigned to this run. All five ran the authenticity-verification protocol; all five passed.

• FDA CDRH, "Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback" — score 8 (live thread +3, uncaptured primary document +3, US AI regulation core +2). The highest-scoring gap this run. FDA — Food and Drug Administration (AI Deployer) carried this development from a lede-only Inside AI Policy report and said so in the page text: "the docket number, the comment deadline and the full list of questions posed are not established here." Verified: fda.gov, on two independent paths (the DHCoE landing page and the FDA Newsroom announcement), both dated August 18, 2026. Docket FDA-2026-N-7874, comments due October 19, 2026. Queued URL-only for a fresh fetch of the paper at ingest, per the QUEUE lane's large-PDF allowance. Verification record: queue/gap-scan/proposed-sources/fda-genai-device-discussion-paper-2026.md. Queued: INGEST-fda-genai-device-discussion-paper.md.

• H.R. 9965, AI Threat Output and Monitoring Incident Containment Act (Introduced in House, 2026-07-27) — score 6. Verified: congress.gov. Saved: Raw Sources/H.R. 9965 - AI Threat Output and Monitoring Incident Containment Act (Introduced in House, 2026-07-27).md. Queued: INGEST-hr-9965-ai-threat-output-monitoring.md.

• S. 5130, Human Dignity and Emerging Technologies Act (Introduced in Senate, 2026-07-23) — score 6. Verified: congress.gov. Saved: Raw Sources/S. 5130 - Human Dignity and Emerging Technologies Act (Introduced in Senate, 2026-07-23).md. Queued: INGEST-s-5130-human-dignity-emerging-technologies.md.

• New York A8833, Understanding Artificial Intelligence Act (Introduced in Assembly, 2025-06-09) — score 6. Verified: nysenate.gov Open Legislation. Saved: Raw Sources/New York A8833 - Understanding Artificial Intelligence Act (Introduced in Assembly, 2025-06-09).md. Queued: INGEST-ny-a8833-understanding-ai-act.md.

• Rhode Island H 8052, Artificial Intelligence — Liability for Injuries (Introduced, 2026-02-27) — score 6. Verified: webserver.rilegislature.gov. Saved: Raw Sources/Rhode Island H 8052 - Artificial Intelligence Liability for Injuries (Introduced, 2026-02-27).md. Queued: INGEST-ri-h8052-ai-liability.md.

Authenticity-verification failures

None. One near-miss is recorded rather than acted on: a search for the August 2026 FDA item returns docket FDA-2026-N-4390, "AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program; Request for Information" (91 FR 23100, published 2026-04-29, comment period extended to 2026-06-29). That is a different instrument on a different subject, already covered on FDA — Food and Drug Administration (AI Deployer). The two are easy to conflate precisely because the coverage of the August item called it a "request for information" — the term that belongs to the April one. Recorded in the verification file so a later run does not repeat the confusion.

Findings worth stating once

• Four of the five queued documents were misdescribed by the coverage that reached the wiki, in the exact classes source-fidelity names. (1) Instrument class. The FDA published a discussion paper and request for feedback, explicitly "not draft or final guidance" and "not intended to propose or implement policy changes"; the digest and FDA — Food and Drug Administration (AI Deployer) call it a request for information. (2) Subject scope. It concerns generative-AI-enabled medical devices, not "the regulation of generative artificial intelligence". (3) Compression. Coverage said the technology would be "shadowed side by side with human clinicians"; the FDA's own words are "non-clinical device benchmarking and clinical confirmation". (4) Date. NY A8833 was introduced June 9, 2025, not 2026 — it surfaced in a July 2026 digest because of a January 2026 re-referral. Each correction is written into the relevant ingest task.
• Two bills queued this run are near-twins, and the differences are the interesting part. NY A8833 and RI H8052 share definitions, a rebuttable presumption that an AI system holds the relevant mental state, an express bar on arguing that AI systems cannot have mental states, and two parallel defenses. They diverge in three ways that change the regime: A8833 reaches only "covered models" above 10^26 operations and $100 million in training cost, where H 8052 reaches any AI system with no threshold; A8833's defenses are affirmative, where H 8052's are absolute and expressly displace common-law and existing statutory liability; and A8833 confines recovery to injuries to a non-user, where H 8052 does not. Whether they descend from a shared model bill is worth establishing from a source and was not asserted.
• H.R. 9965 contradicts itself. Its long title directs "the National Nuclear Security Administration through the Secretary of Energy"; section 2(a) directs the Secretary of Energy "acting through the National Laboratories", and the NNSA appears nowhere in the operative text. Recorded in the ingest task rather than resolved by inference.
• RI H 8052 uses a term it never defines. The liability rule at (b) attaches to "artificial intelligence system"; both defenses at (d) attach to "covered model or covered model derivative", which subsection (a) does not define. Read literally the defenses may not reach the conduct the rule covers. Recorded as what the text says; the New York definition was deliberately not imported to fill it.
• The ## Primary documents section is missing from every digest since it was added. Schema v4.15 (2026-08-20) added step 2.5(f) to the developments-brief routine so digests record the URL of the document behind a story. Neither 2026-08-20 digest has the section. _automation/routines/ is the source of truth for the text, but the copy that runs lives in the Claude Desktop UI and must be re-pasted by hand — the drift the v4.14 changelog warned nothing local can detect. Until it is re-pasted, gap type 9's precise lane stays empty and the work falls to the keyword pre-filter, which is exactly the weaker instrument v4.15 was written to replace. This is the one item on this list that needs the curator rather than a later scan.
• entities/cisa.md and government/cisa.md both exist. CLAUDE.md's no-duplication rule allows an organization in exactly one folder. CISA is a deployer, which points to government/; whether the entities/ page covers a genuinely distinct regulatory role is a curation call, so it is recorded rather than merged. New this run, found while wiring DHS — Department of Homeland Security (AI Deployer).
• The escaped-pipe scanner artifact is fixed. Recorded as an open finding on eight consecutive runs through 2026-08-17. bin/lint-scan.py now resolves [[page\|alias]] before counting and says so in the report header. No run should re-derive it.

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

• The remaining primary-document backlog (score 5–6). Lane 1: the White House release of 2026-07 (whitehouse.gov) — deferred deliberately, because the needs-review note flags that a release is not always the primary text it announces and the instrument must be identified before queueing. Lane 2 possibles: the Alphabet Q2 2026 earnings release, and the ANI v. OpenAI Delhi High Court order of 2026-07-24, where the cited host is a blog reproduction and the authenticity protocol prefers the court's own docket.
• Slice-8 thin anchors not reached (score 3–4 each). government/dot (397 words, in-degree 6, last_updated: 2026-06-06); government/cbp (199 words, in-degree 5, confidence: high on 2 sources — the thinnest high-confidence page in the slice, and now under-supported relative to the DHS expansion that cites CBP use cases it does not carry). Thin litigation pages at in-degree 3–4: cox-v-sony (261 words, 1 source), spacexai-v-bonta (206, 1), acls-v-neh (262, confidence: low, 1), sony-v-udio (271, low, 2), chatrie-v-united-states (353, 2).
• Live-thread items named in the window with no page (score 3 each). Companies: Marvell (the $12.18 billion Google warrant, carried across two digests, with full terms in the August 20 morning file), Vercel, Slack, CrowdStrike, Casper Studios. Note a name collision to resolve before any page is created: the venture firm Elia Zaitsev is launching is called Cognition, and companies/cognition.md already exists for a different company. Also: the Waymo in-vehicle compute disclosure (custom 5nm ASIC, 1,000 TOPS, named suppliers), OpenAI's Private Safety Processing, Anthropic's planned data-retention change, the Irregular postmortem criticism, New York's FutureWorks Commission, and the August 18 Pew Research Center report on American attitudes to AI — the last of these is a foundational-source candidate, and sources/pew-americans-and-ai-2026 should be checked for whether it already is this report before one is queued.
• The 24 referenced-2+ broken targets, unchanged. Twelve are author entity pages implied by sources/ pages (entities/kevin-klyman, entities/lee-anne-fennell, entities/michael-j-d-vermeer, entities/sacha-altay, entities/joseph-bernstein, entities/clayton-christensen, entities/samuel-weinbach, entities/ilhan-scheer, entities/dina-powell-mccormick, entities/adam-smith, and others), each at in-degree 2. Recorded for a seventh consecutive run. This is structural, not transient, per the skill's three-run rule: a batch pass over the author-page set is the right instrument, not two or three per run, and recording it again without routing it there is itself the finding.
• Institution shorthand targets (score 1–2), unchanged. entities/carnegie-mellon, entities/santa-fe-institute, entities/bank-of-england, entities/ecb, entities/federal-reserve, entities/uw. Sixth consecutive run. uw is an abbreviation better rewritten to a full name in the citing prose than created as a page; that rewrite is lint's lane.
• Concept targets named but pageless (score 2 each), unchanged. concepts/inverse-cooking-problem and concepts/inverse-trust-problem (held under needs-review/2026-07-10-inverse-cooking-trust-coined-terms.md), concepts/a-vision-of-democratic-ai, concepts/data-broker-regulation, concepts/talent-flow-china-us, concepts/economic-possibilities-for-artificial-intelligence, concepts/ai-and-language-models, concepts/algorithmic-decisionmaking (seventh consecutive run; a disposition note exists at needs-review/2026-08-19-algorithmic-decisionmaking-vs-admt.md).
• The 29 date-stamped headers in lint-report.md (score 1). Unchanged for five cycles. Some are legitimate procedural histories, which the report does not distinguish; that remains the finding.

Checks run before closing

• No broken outbound links introduced. All wikilinks on the two new pages, the expanded page and the five queue files were checked against the page set after writing; zero unresolved. Three placeholder [[sources/…]] links written during drafting were caught and rewritten as plain text before the run closed, and one wrong slug (legislation/ny-raise-act → legislation/new-york-raise-act) was corrected.
• Broken-link count unchanged at 235 distinct / 259 refs. Expected: none of the actioned gaps was on the referenced-2+ list, because all were dangling prose references, live-thread items or primary documents rather than broken wikilinks. The same pattern held on 2026-08-20.
• Style scan clean. Both new pages and the expanded page checked against the banned-header, hyperbole, self-reference, standing-Predictions and date-header lists. Two apparent hyperbole hits were inspected and dismissed: "soars" appears only inside a Yahoo Finance URL slug, and "essential" only inside a direct quotation from Minister Anika Wells, which house style §10 permits.
• Revision fidelity verified mechanically, not by eye. See the DHS — Department of Homeland Security (AI Deployer) entry above.
• Confidence set honestly. Both new pages are confidence: medium; neither reached high, and the three-independent-source bar was not met for either. DHS — Department of Homeland Security (AI Deployer) retains the high it already carried, now reinforced rather than assumed.
• Cap respected. Eight gaps actioned against a 6–10 cap; thirty-four candidates found. Ten web pulls and four searches, no crawls.

One-line summary

Two pages created and one substantially expanded; five foundational sources verified and queued, clearing four of the five lane-1 primary-document items the previous run assigned to this one — and the curator's attention is needed on one thing only: the developments-brief routine in the Claude Desktop UI has not been re-pasted since schema v4.15, so no digest yet carries the ## Primary documents section that change added.