AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-04

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

Contents

Scanned

Recent window: 6 New Developments Log/ files dated 2026-09-01 to 2026-09-03, and 93 content pages with a mtime inside 48 hours. Rotation slice: 8 — litigation/ + standards/ + government/ (92 pages).

Tooling run: bin/lint-scan.py (212 distinct broken targets across 212 references, 0 referenced 2+ times), bin/primary-doc-scan.py --days 3 (6 digests; 5 cited primary-document URLs, 1 with no raw file, 1 possible match under another address; 4 keyword pre-filter hits), bin/devlog-backlog.py (270 folded, 1 unprocessed).

Gaps actioned (7 of 22 found)

New pages created (live)

  • ARC Prize Foundation — score 5 (gap type 4 +3, safety/evaluations core area +2). A foundational ARC Prize source is ingested (ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence) and two more sit in Raw Sources/ (the ARC-AGI-2 and ARC-AGI-3 reports), while "ARC Prize" appears on 12 content pages including index.md, three model pages, three concept pages and two sources/ pages — with no organization page. Alias resolution found nothing: entities/ holds no arc* file, and concepts/arc-agi is the benchmark, not the body that maintains it.

1,020 words, confidence: medium, 5 sources. Scoped deliberately to the organization rather than the benchmarks, which ARC-AGI already covers at length: origin as a June 2024 Kaggle competition (1,400+ teams, ~17,000 entries, $600,000 grand prize unclaimed, $75,000 to three papers and $50,000 to five teams), constitution as a foundation in 2025 with a $725,000+ prize pool on ARC-AGI-2, the nine named staff and board members, the stated mission and its falsification criterion, and the third-party evaluation role including the December 2024 o3 verification and the September 3, 2026 Astra evaluation with its harness caveat. Two open questions recorded rather than papered over: the foundation publishes no funder list or audited financials that this run located, and whether the Astra evaluation's harness distinction will be applied retrospectively is not stated.

Pages expanded (live)

  • Massachusetts economic development bill — AI safety provisions (2026) — score 8 (live thread +3, uncaptured primary document +3, US state AI regulation core +2). Created by the nightly fold on 2026-09-03 from a single paywalled trade report, with sources_count: 1, no bill number, and the explicit statement that "no primary text has been retrieved." 490 → 1,082 words, sources_count 1 → 5.

Added from malegislature.gov: the vehicle is H.5576 of the 194th General Court, not the standalone S.2630; the Senate Ways and Means 16–0 redraft vote of July 16, 2026 and Senate passage on July 23; the Attorney General civil-enforcement mechanism; the "large frontier developer" framing the Senate itself uses; the $575.4 million in authorizations including $75 million for AI and $25 million for robotics; and the three companion technology amendments (94 on AI-generated CSAM, 118 on weaponized robots, 374 on crypto kiosks). Four open questions replaced the original three, all sharpened rather than answered. [5 new (Source: URL) cites; confidence held at medium.]

  • NetChoice v. Bonta (CAADCA litigation) — score 7 (dangling foundational reference +3, in-degree 16 +2, core area +2), the top thin anchor in slice 8 by in-degree. 765 → 1,268 words, sources_count 2 → 6, last_updated 2026-06-06 → 2026-09-04.

Four defects fixed. (1) The page carried zero citations of any kind — every claim, including the two opinions' holdings, was unsourced. (2) confidence: high on sources_count: 2, which the decay table does not permit; lowered to medium. (3) A duplicate status: key in the frontmatter — the base-schema page status and the litigation case status both written to status, so the second silently overrode the first; the case status is now case_status, matching the convention on Sony Music v. Udio (second suit, 2026) and the rest of the folder. (4) **plaintiff:** NetChoice (planned entity page) pointed at a page that has existed for some time; rewritten as a typed litigates: edge to NetChoice.

Material added: the docket (No. 25-2366), the district-court number and judge (5:22-cv-08861-BLF, Judge Beth Labson Freeman), the argument date (January 14, 2026), the panel (M. Smith, Bennett, Johnstone) and author, publication status and length, all read off the slip opinion; the opinion's own disposition line; the provision-by-provision mapping of which subsections of Cal. Civ. Code § 1798.99.31 were affirmed on vagueness grounds and which were lifted; and Holland & Knight's March 2026 note that the CAADCA may remain unenforceable for some time if the state again stays enforcement on remand. The status line is dated to September 4, 2026. [5 new (Source: URL) cites.]

  • CBP — Customs and Border Protection (AI Deployer) — score 3 (US government AI deployment core +2, in-degree 6–9 +1). At 199 words it was the thinnest page in slice 8 against its in-degree of 8, confidence: high on sources_count: 2, last_updated: 2026-06-06, with no citations and a stray paragraph about Google's consumer age-estimation rollout sitting unattached to any CBP fact. 199 → 1,105 words, sources_count 2 → 7, confidence lowered high → medium.

The page described two AI uses. The DHS AI use-case inventory for CBP, dated July 15, 2026, records dozens. Added a table of 19 named use cases with their DHS identifiers and deployment status — autonomous surveillance towers, KESTREL landing prediction, vessel and item-of-interest detection, six cargo and trade-risk models, passport anomaly detection, the Dataminr and Babel open-source screening tools, Airship Outpost, and Mobile Fortify — plus the traveller-verification family and CBP Link/TVS inside CBP One. Added the procurement record: the Clearview AI contract giving access to a database of more than 60 billion images for "tactical targeting" and "counter-network analysis," and the Brennan Center's tracking of DHS-wide facial-recognition spend (at least $47.188 million 2021–2025; ~$40 million to NEC through OBIM, ~$7.188 million to Clearview across ICE and CBP), explicitly labelled department-wide rather than CBP-specific. Added the reliability record: Wired's February 2026 report that the CBP and ICE face-recognition application cannot verify identity, and the removal of DHS Directive 026-11 from the DHS website. Every original fact and framing survives, including the Altana contract, the MIT Technology Review item and the Google age-estimation paragraph. [6 new (Source: URL) cites.]

Graph wiring

Six additive wikilink edits across five existing pages. Each wraps a link around a name already in the prose or corrects a bare link; no sentence was rewritten and no fact or citation touched.

Zero — ninth consecutive run. bin/lint-scan.py reports 212 distinct broken targets across 212 references, 0 of them referenced 2+ times, before and after this run. Step 3 restricts candidate generation to the referenced-2+ lane and directs that the single-reference long tail not be re-audited. Reported without re-auditing.

Queued — foundational sources

  • David Feith, A New Counterintelligence Map for the AI Age, prepared statement before the House Permanent Select Committee on Intelligence (2026-09-02) — score 8 (live thread +3, uncaptured primary document +3, national-security core +2). Gap type 9: the 2026-09-03 08:05 digest carried the hearing item entirely through Inside AI Policy, with no primary-document URL. Verified: Hudson Institute's own media bucket, HTTP 200, 10-page PDF retrieved complete; the sentence Inside AI Policy quotes appears verbatim; the document self-identifies the committee, hearing title, chairman and ranking member. Saved in full: Raw Sources/David Feith - A New Counterintelligence Map for the AI Age (HPSCI Prepared Statement, 2026-09-02).md. Queued: INGEST-feith-counterintelligence-map-ai-age-2026.md.
  • Ninth Circuit, NetChoice, LLC v. Bonta, No. 25-2366 (2026-03-12) — score 7 (dangling foundational reference +3, in-degree 16 +2, core +2). Gap type 1: the opinion the expanded litigation page rests on, with no Raw Sources/ file and no sources/ page. Verified: cdn.ca9.uscourts.gov, the court's own opinion datastore, HTTP 200, 46 pages, FOR PUBLICATION; corroborated by the Justia docket entry and four independent practitioner analyses. Queued URL-only — see the verification failure section below for why no raw was saved. Queued: INGEST-ca9-netchoice-v-bonta-opinion-2026.md.
  • Massachusetts H.5576, economic development bill (AI provisions), Senate-passed 2026-07-23 — score 8, same basis as the page expansion above. Verified: malegislature.gov, two pages HTTP 200; the vehicle identified through the Senate press release's own link to the H.5576 committee vote. Queued URL-only (a $575.4 million bond bill; the engrossed AI article should be extracted fresh at ingest rather than transcribed from a scrape buffer). Queued: INGEST-ma-h5576-economic-development-ai-2026.md.

Three verification records written to Wiki/_meta/queue/gap-scan/proposed-sources/.

Source-fidelity findings carried into the queue tasks

Four divergences between the record and the digests or the existing pages, recorded in the INGEST- tasks rather than fixed silently.

  1. The Massachusetts four-month requirement is a floor amendment, not the underlying draft. The digest, and the page the fold built from it, treat in-depth independent safety testing every four months as the bill's core requirement. It entered by Amendment 471, "Frontier AI Risk Reports and Independent Verification," adopted during the July 22–23 Senate debate, which added both third-party compliance audits and the four-monthly testing. Anthropic's support and Google's and OpenAI's opposition are therefore a split over an amendment, not over the Senate Ways and Means redraft. The page now says so.
  1. The Feith testimony is dated a day before the hearing, and is not about data centres. The digest dates the item September 3, 2026 and frames it as urging members to treat data centres as a cybersecurity target. The document is captioned September 2, 2026, and data centres are section 3 of eight. Its stated argument is that the United States needs a counterintelligence map for the AI age, organised around five ecosystems — sensitive sites and drone exposure, the grid and data-centre hardware, frontier AI labs, advanced compute and the semiconductor supply chain, and startups crossing a "counterintelligence threshold" — with five lettered recommendations. Both dates and the full scope are carried into the ingest task.
  1. The Feith statement carries a dozen figures the digest does not, each attributed inside the document to a third party and listed item by item in the queue task: the July 22 Northern Virginia fault that moved roughly 3,800 MW of data-centre load to backup generation with voltage and frequency swings from Virginia to Chicago; Operation Spiderweb's 117 drones against four airfields on June 1, 2025; roughly 17 nights of drones over Langley AFB in December 2023; Strider's count of more than 2,700 Chinese papers on grid vulnerabilities; the Institute for Progress estimate of Huawei at one to two percent of U.S. advanced-AI-compute production; and the January conviction of former Google engineer Leon Ding over more than 2,000 pages of confidential AI information.
  1. The malegislature.gov press-release page displays a publication date one day after its own dateline — "July 24, 2026" against "(BOSTON—7/23/2026)". The send-vs-event distinction the source-fidelity skill exists to catch, in miniature. The page uses July 23 for passage.

Authenticity-verification failures

None. Four documents sought, four retrieved from a canonical host on the first attempt — the Hudson Institute's own media bucket for the testimony, the Ninth Circuit's opinion datastore for the slip opinion, malegislature.gov for the bill record, arcprize.org and dhs.gov for the two page builds. No rejections, no secondary-copy fallbacks. The trade-press and Justia copies of the CA9 opinion were located and set aside in favour of the court's own file, and the netchoice.org-hosted copy was not used.

One capture was withheld from Raw Sources/ rather than saved, which is a different outcome from a verification failure and is recorded as such.

The Ninth Circuit opinion is authentic and retrieved complete, but its 80,465-character payload landed on a host path this sandbox cannot slice — the limitation recorded on 2026-08-26 and 2026-08-31. A targeted jsonOptions extraction was run in its place, and its provision-level output conflicts with the published practitioner accounts: the extraction reports the injunction affirmed as to Cal. Civ. Code §§ 1798.99.31(a)(1) and (a)(2) and vacated as to §§ (a)(5), 1798.99.30(b)(1) and 1798.99.31(b)(7), while Holland & Knight's provision table — consistent with Cooley's account — reports it affirmed on vagueness grounds as to §§ 1798.99.31(b)(1), (b)(3), (b)(4) and (b)(7) and lifted as to §§ (a)(6), (a)(8), (a)(10) and (b)(8). They cannot both be right and the extraction is the model-mediated one.

The consequence was applied rather than noted: the extraction's section mapping was not written to the page. Only the structural facts an extraction reads off a caption page and that were corroborated externally — docket, D.C. number, dates, panel, author, disposition, publication status, page count — were used from the opinion. The provision-level detail on NetChoice v. Bonta (CAADCA litigation) is sourced to the practitioner analyses as (Source: URL) and carries a > **Provenance note:** saying so, and the ingest task instructs the fold to replace both with the opinion's own text and correct the page if the analyses are wrong.

A second, smaller model-mediated figure is flagged on its own page: the count of 69 CBP entries and 43 marked Deployed in the DHS inventory came from a structured extraction rather than a full read of the page. The 19 named use cases and their identifiers are individually verifiable at the cited URL; the aggregates are not, and the page presents them as what a structured read returned.

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

  • Live-thread items owned by the nightly fold, not by this skill (score 3 each). The 2026-09-03 22:05 digest is unprocessed at scan time (unprocessed_devlog=1) — the third consecutive report of an unprocessed evening digest. Its page-level work belongs to developments-log: the GPT-6 Astra release and the ARC Prize evaluation of it, the Sanders–Casar Ban Artificial Superintelligence Act, the Zuckerberg–Trump call on a national AI regulator, the Pentagon's maintained Anthropic supply-chain designation, the G20 light-touch framework, Crusoe's $30bn round, the Thinking Machines Lab talks, Meta's retreat from AI-usage performance criteria, and the LAUSD device-level block. All were left alone.
  • legislation/caadca — the statute itself has no page (score 5), new this run and the most substantive miss. California AB 2273 (2022), codified at Cal. Civ. Code §§ 1798.99.28–1798.99.40, is reached across the vault entirely through the litigation challenging it. NetChoice v. Bonta (CAADCA litigation) enumerates eight of its requirements and its penalty structure; Age Verification and AI and Children both turn on it. Alias resolution found nothing under caadca, age-appropriate or ab-2273 in legislation/. Deliberately not created this run: a statute page built from the record of the case attacking it inherits that framing, and the right order is to build it from the codified text. Recorded in INGEST-ca9-netchoice-v-bonta-opinion-2026.md so it is not lost.
  • entities/david-feith (score 4) and entities/hudson-institute (score 3) — new this run. Feith is named on six content pages and is now the author of a captured foundational source; Hudson on two. Both surfaced while verifying the testimony and both lost to the three higher-scoring captures. Feith is the stronger of the two and will likely clear the cap next run.
  • Slice-8 thin anchors, the remaining four. The slice is in better shape than the concepts/ and entities/ slices: the standard diagnostic (in-degree ≥6 with sources_count ≤ 2, missing sources_count, no ## Relationships, or under 350 words) returns 7 of 92 pages, against 84 of 198 in slice 4 and 24 of 126 in slice 6, and every page in the slice has a Relationships section. Worked: netchoice-v-bonta, cbp. Remaining, by score: Getty Images v. Stability AI (US) (in-degree 9, 562 words, sources_count: 2); ISO/IEC 42005 — AI Impact Assessment (8, 642 words, sources_count: 2, confidence: high, last_updated: 2026-06-06); Tremblay v. OpenAI (8, 619 words, sources_count: 2); GEMA v. Suno (Munich Regional Court) (8, 302 words, sources_count: 1); Sony Music v. Udio (second suit, 2026) (7, 271 words, confidence: low, last_status_check: 2026-07-22 — 44 days on a case the schema puts in the 3-month fast-decay band, and its court field still reads "reported; venue follows the original S.D.N.Y. case"). The two music cases are the ones that most need a docket check rather than a web pass; the CourtListener MCP is unauthenticated in this session, which is why neither was actioned.
  • entities/adam-thierer (score 3) — seventh consecutive deferral, structural. Unchanged; two mentions.
  • models/deepseek-v3 — fourteenth run, structural. In-degree 49 on 608 words, routed to model-robustness-check on 2026-08-22, unaddressed.
  • Date-stamped headers in lint-report.md (score 1) — nineteenth cycle, structural, holding at 26. Six self-references also appear in this run's lint output, unchanged.
  • Carried unchanged. Marvell (no company page, carried since 2026-08-21); California AB 1856 (text not located); the Texas OAG / Meta Webster–Mahoney exchange; the draft executive order on a frontier-developer self-regulatory organization — which the 2026-09-03 22:05 digest advances but does not make actionable, since the reporting is now that Zuckerberg opposed it in a private call and the administration is weighing two models with no text published either way; the AI Evaluator Forum / AEF-1 standard (score 4); ExploitBench flag-level disclosure (score 3); the "Carolina Principles"; OpenAI's "Our agreement with the Department of War"; the House E&C markup memo (HMTG_119_IF_17_20260901_SD_002, the run's single lane-1 hit, a committee staff memo and therefore supporting); the METR/Redwood September 1 Hugging Face analysis; the OpenAI letter to Reps. Casar and Matsui of September 2 (no public copy); the New York City Public Schools generative-AI policy; and the Agentic Commerce Protocol / Model Context Protocol placement question routed to the SIIA ingest.

Dedupe drops (all correct)

Five candidates were dropped as done or in flight rather than scored. The Brinkema ad-tech remedy memorandum and the Amazon rehearing en banc petition each carry an open INGEST- task. The FTC v. Amazon complaint, the SB Energy S-1 and the SIIA amicus — three of the four keyword pre-filter hits from primary-doc-scan.py — already have sources/ pages or open queue tasks. The Anthropic August 2026 Risk Report, the scan's "possible match under another address" hit, is held under its canonical landing page rather than the CDN path the digest cited; this is the known false-positive shape recorded in the v4.15 changelog, not a gap.

Run integrity

  • No broken outbound links introduced. All nine touched pages were resolved against the full page set after writing: zero unresolved on every one.
  • Style scan clean, first pass. All nine checked programmatically against the banned-header, hyperbole, self-reference, standing-Predictions and date-led-header patterns: zero hits on each.
  • Revision fidelity verified mechanically, not asserted. For all three expanded pages: citations 0→5, 0→6 and 1→6 with none dropped; wikilinks 6→10, 4→7 and 10→10 with none dropped; and a token-level diff of every number, percentage, dollar figure and year found zero numeric facts dropped. Word counts 765→1,268, 199→1,105 and 490→1,082.
  • Confidence set honestly, and lowered twice. The one new page is confidence: medium on five sources. NetChoice v. Bonta (CAADCA litigation) and CBP — Customs and Border Protection (AI Deployer) were both reduced from high to medium, because each carried high on sources_count: 2 in violation of the decay table and each rests in part on claims this run did not re-verify. Neither was raised to compensate for the material added.
  • A frontmatter bug was fixed rather than worked around. NetChoice v. Bonta (CAADCA litigation) carried two status: keys, the base page status and the litigation case status; in YAML the second silently overrides the first. The case status moved to case_status, matching the rest of litigation/.
  • Cap respected. Seven gaps actioned against a 6–10 cap; 22 candidates found. Eight scrapes and four searches, no crawls, no failures.
  • Dashboard rebuilt — see the audit entry.

One-line summary

One new page (ARC Prize Foundation) and three expansions (Massachusetts economic development bill — AI safety provisions (2026), NetChoice v. Bonta (CAADCA litigation), CBP — Customs and Border Protection (AI Deployer)) went in live; three foundational sources are queued for review, one of them saved in full; the Ninth Circuit CAADCA opinion's provision-level mapping is contested between a structured extraction and four practitioner accounts and was deliberately kept off the page pending ingest — that is the item most needing the curator's eye.

Categories: Briefings | briefing | gap-scan