Gap Scan — 2026-08-28

Scanned

Recent window: 4 New Developments Log/ files (2026-08-26 08:05 and 22:04; 2026-08-27 08:10 and 22:05), 68 wiki pages with last_updated in the last 48 hours. Rotation slice: 1 — concepts/ A–F (185 pages).

bin/lint-scan.py and bin/primary-doc-scan.py --days 3 both run; the lint report is the authority for broken links, per the standing rule against hand-rolled scans.

Broken-link lane is empty for a second consecutive run. 214 distinct broken targets, 0 referenced 2+ times, 214 referenced once. Candidate generation is restricted to the referenced-2+ list, so the lane produced no candidates. The long tail was not re-audited.

Primary-document scan: 6 digests, 11 cited primary-document URLs, 6 with no Raw Sources/ file, 2 possible matches under another address, 3 keyword pre-filter hits in the named lane.

24 candidates found; 6 actioned.

Gaps actioned (6 of 24 found)

New pages created (live)

None. No candidate this run was a missing-page gap that passed the quality gate — the run's weight sat on one live litigation thread with newly available primary documents and on the rotation slice's two worst-supported high-in-degree anchors. Recorded explicitly rather than left as an empty heading.

Pages expanded (live)

• Anthropic v. United States (Pentagon ban challenge) — score 7 (live thread +3, in-degree 40 → +2, frontier-lab governance and US policy core +2). The case reached final judgment on 2026-08-27 and the page had not been touched since 2026-07-31. Expanded from the court's own filings rather than from press accounts, with the primary documents queued and a > **Provenance note:** recording that citations should upgrade once they are ingested. Added a ### Summary judgment and final relief (August 27, 2026) subsection covering the administrative record ("the record is slim" — a four-page Michael memorandum post-dating two of the three challenged actions), the government's abandonment of its backdoor-access theory and its concession that Anthropic's technology is "no riskier to the national security than any other 'black box' artificial intelligence model," the residual "trust" ground, the three contemporaneous facts the court read against pretext (the Defense Production Act proposal, the "we are very close here" contract pursuit, the continuing Mythos collaboration discussions), a five-row count-by-count disposition table, the three distinct remedies entered, and the order's own stated limits (¶ 14) and retention of jurisdiction (¶ 15). Four frontmatter corrections the primary record settles: the court is N.D. Cal., No. 3:26-cv-01996-RFL where the infobox read "exact district not confirmed"; the lead defendant is the U.S. Department of War; the agency defendant list is far wider than "United States of America; Department of Defense"; and case_status moves active → judgment-plaintiff. Added an ## Open questions section (appeal status, fate of the D.C. Circuit petition, the July 31 Lin remark). 1,544 → 2,858 words, sources_count 6 → 9, confidence: high held, not raised — the rating was already high and is now reinforced by primary documents within the fast-decay window for active-litigation status. last_status_check 2026-07-02 → 2026-08-27.
• AI and Civil Liberties — score 4 (in-degree 32 → +2, US regulation/policy core +2). The worst-supported page in the rotation slice on any reading: 458 words, sources_count: 0, not one citation anywhere in the body, and 83 days stale, while 32 pages link to it. Added ## The developer as First Amendment claimant, covering the line of civil-liberties litigation that runs through the AI developer rather than the person subject to the system, built on the 2026-08-27 order and opinion; and ## Private surveillance networks, covering the ACLU's August 2026 analysis of Flock Safety's announced changes — the 30-day-to-one-week retention change and the "Evidence Mode" exception that governs whether it means anything, the previously circumvented search-purpose requirement, the unevaluated audit-assistance tool, the 120,000 readers across at least 6,000 cities, and the Institute for Justice database of 100+ abuse incidents. The ACLU/CDT amicus position is presented as a position — that a developer's use restrictions operate as a de facto privacy safeguard where statute is absent — with the organizations' own framing of it as second-best to legislation. 458 → 1,148 words, 0 → 5 sources. confidence: medium held: the page is now cited but remains an umbrella whose substance sits on the pages it points to.
• AI Diffusion — score 4 (in-degree 37 → +2, compute/competition core +2). 574 words, sources_count: 1, 83 days stale, and carrying the concept almost entirely as an argument from a single Foreign Affairs essay with no measurement of the thing being argued about. Added ## Measurement and ## Barriers in lower-income countries, from four independent sources: the definitional problem (Brookings, including the U.S. Census February 2026 figure of ~17.5% of U.S. businesses and the November 2025 question rewrite that makes the series discontinuous); Microsoft's Q1 2026 Global AI Diffusion Report (17.8% of the global working-age population, 26 economies above 30%, UAE 70.1%, US 21st at 31.3%, North 27.5% against South 15.4% and a 12.1-point gap widening at roughly twice the South's rate of gain), with the company's own caveat that no single metric captures adoption; the underlying population-normalized metric paper (147 countries against Stanford HAI's 36, and its finding that the binding constraint is internet access rather than demand); and the World Bank's May 2026 analysis of the material constraints — ChatGPT visits per internet user ~50× high-income to low-income, a 5GB package taking nearly half of the post-food household budget, under 5% of low-income-country populations with basic digital skills in 2023, English at nearly half of global URLs, and 77% / 23% / <0.1% of global data-center capacity by income group. 574 → 1,436 words, 1 → 6 sources, confidence: medium held.

Link aliases fixed

Zero. Sixth consecutive run. The finding is settled and was not re-audited: the referenced-2+ lane is empty, and the 214 remaining targets are author surnames, institution shorthand and deliberate forward references.

Queued — foundational sources

• Order of Final Relief, Anthropic PBC v. U.S. Department of War, Dkt. No. 251 (N.D. Cal., 2026-08-27) — score 8 (live thread +3, uncaptured primary document +3, core +2). The highest-scoring gap this run and the only lane-1 hit from primary-doc-scan.py not already queued. Verified on the CourtListener RECAP docket gov.uscourts.cand.465515: caption, case no. 3:26-cv-01996-RFL, Filed 08/27/26 stamp and Judge Lin's signature block present; the companion opinion cross-references it by number; disposition corroborated by the WSJ and Yahoo/AP reports the digest cites. Captured in full (4 of 4 pages). Saved: Raw Sources/Anthropic PBC v. U.S. Department of War - Order of Final Relief (N.D. Cal., 2026-08-27).md. Queued: INGEST-anthropic-v-dow-final-relief-order.md.
• Order on Cross Motions for Summary Judgment, same case, Dkt. No. 250 (59 pages) — score 8. Found by following the citation rather than by the scan, and this is the run's substantive finding (below). Verified from three pages fetched live; both sentences quoted in press coverage confirmed verbatim in the introduction. Queued URL-only under the large-PDF allowance, with the table of contents recorded in the ingest task so the eventual capture can be checked for completeness, and an explicit note that a three-page verification read is not a capture. Queued: INGEST-anthropic-v-dow-summary-judgment-opinion.md. Verification record for both documents: queue/gap-scan/proposed-sources/anthropic-v-dow-summary-judgment-2026.md.
• "A call for collective action on cyber defense," open letter, 128 signatories (2026-08-27) — score 8 (live thread +3, uncaptured primary document +3, safety/national-security core +2). Named-lane hit promoted after reading the item: the letter is the instrument the digest item is about, the digest records it as "retrieved and read in full," and yet a search for "collective cyberdefense" and "collective cyber" across the content folders returned zero hits. Verified: fetched live from openai.com with cache bypassed, HTTP 200; corroborated by Axios of the same date. The signatory block is client-rendered and absent from a default fetch — a rendered fetch recovered it, and the full list is in the raw file. Saved: Raw Sources/A Call for Collective Action on Cyber Defense (open letter, 2026-08-27).md. Queued: INGEST-collective-cyberdefense-letter-2026.md.

None of the three was ingested. Folding a primary source into the graph stays a curated step.

Authenticity-verification failures

None. Three documents pulled, three verified, zero rejected.

Findings recorded against the record

Two, both of the kind the schema's primary-document machinery exists to surface, and both invisible to a scan that stops at "was a URL cited."

1. The 2026-08-27 22:05 digest names the right document and cites the wrong one. Its ## Primary documents block lists …465515.251.0_1.pdf under the label "Opinion and order of Judge Rita F. Lin, N.D. Cal." That URL is the four-page order of final relief. The 59-page opinion the digest quotes — the source of both "the empty invocation of national security is not a blank check to punish and retaliate against government critics" and "[n]either the Constitution nor the federal statute invoked by Defendants allows them to impose sweeping penalties based principally on Anthropic's critique of the Administration's views" — is Dkt. No. 250, at a different address. bin/primary-doc-scan.py reported the citation as a clean lane-1 hit because a URL was present and no raw file matched; nothing mechanical could catch that the URL and the label described different documents. Both are now queued. The general case is worth noting: the lane-1 check verifies that a cited document is captured, not that the cited document is the one the item rests on.
2. The signatory list of the cyber-defense letter is client-rendered. A default fetch of the letter returns the full statement and an empty ## Supporting organizations heading. The count is 128 organizations, against the "more than 100 other companies" that press coverage and the digest carry, and the composition — 128 named firms spanning frontier labs, hyperscalers, chipmakers, the security industry, major banks and payment networks, insurers and the Big Four — is the letter's substance. Eight signatories have wiki pages that do not record it. A capture taken without a rendered fetch would have looked complete and lost the list.

A third, smaller correction is carried into the ingest task rather than acted on here: Anthropic v. United States (Pentagon ban challenge) dates the supply-chain-risk designation to March 5, 2026 on subscription reporting, while the opinion states the designations were made "On February 27 and March 3, 2026." Recorded on the page as a discrepancy, not resolved, pending retrieval of the underlying filings.

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

• CISA advisory AA26-231A on the U.S. water-systems intrusion (score 5). Named in the 2026-08-27 22:05 digest's ## Primary documents block as the incident behind the cyber-defense letter. Raw Sources/ holds no CISA advisory of any kind; the wiki's only CISA raw file is a 2026-02-05 Grassley letter to the agency. Six content pages touch water-system intrusions. This is the clearest remaining lane-1 gap and the first candidate next run.
• Exponential View, "State of the AI Economy 2026" (score 3). Reported by primary-doc-scan.py as a "possible match under another address" — the state the scanner added specifically to avoid a false positive. Needs a human-eye check of whether the vault already holds this report under a different URL before it is either queued or dropped.
• Louisiana Attorney General's certified petitions against Snap and Discord (score 3), two cited URLs, examined and deferred rather than actioned. Both are lane-1 hits with no raw file. Against them: the 2026-08-27 22:05 digest lists the suits among twelve items dropped for space, so no wiki claim rests on them; "Louisiana Attorney General" and "Murrill" return zero hits across the content folders; and the subject is platform child-safety rather than AI, which is the reasoning applied to companies/tiktok on 2026-08-25. Recorded so the two URLs are not re-raised as an oversight, but they remain uncaptured.
• Microsoft AI Economy Institute, Global AI Diffusion Report Q1 2026 (score 4) — newly surfaced by this run's own work. Folded into AI Diffusion as a supporting source under the expand lane's rules, but an institutional quarterly report with an original measurement methodology is foundational under the classification rules, and the companion metric paper is a preprint the wiki names nowhere. Queue both next run. The same applies to the World Bank's May 2026 Atlas chapter.
• Live-thread items owned by the nightly fold, not by this skill (score 3 each). The 2026-08-27 22:05 digest is unprocessed (unprocessed_devlog=1), so developments-log still owns: the European Commission's first use of its AI Act enforcement powers during the week of August 24, ordering information from leading model developers on cybersecurity, safety and copyright compliance — EU AI Act (Regulation 2024/1689) carries Virkkunen only at May and June 2026 and has nothing on this; Epoch AI's August 27 revenue analysis (OpenAI above a $40B run rate, Anthropic $1B to $9B across 2025 and a reported $65B by end-July 2026, the pair from $30B to $105B combined, with the authors' note that the two book cloud-platform revenue differently), a Snapshot fold; and Nvidia's pause of revenue-sharing deals with AI cloud providers.
• Four slice-1 thin anchors below the cap (scores 4 and 3). Compressed 21st Century (in-degree 41, 728 words, sources_count: 2) and AI Divides (Literacy / Occupational / Ethico-Philosophical) (in-degree 40, 1,010 words, 2 sources) are the two highest-in-degree pages in the slice and were beaten only by having more words than the two actioned. Then Agent Autonomy Spectrum (5 Levels) (35, 821 words, 2 sources) and Alignment Assemblies and Collective Constitutional AI (28, 627 words, 1 source, last_updated: 2026-06-06). Also California Effect (26 in-degree, 456 words, sources_count: 0) — the slice's second page carrying zero citations, and it should be read alongside the actioned one.
• Twenty-one source-robustness-check candidates across slices 11–13 — fifth consecutive report, structural. Unchanged, led by California SB 53 — Transparency in Frontier AI Act (in-degree 254) and TAKE IT DOWN Act — Source Summary (97). Named structural on 2026-08-27 and the position has not moved: gap-identifier's expand lane is for non-source anchors and cannot absorb this, so the skill needs its own scheduled pass. Stated once, not re-listed.
• The 28 date-stamped headers in lint-report.md (score 1) — twelfth cycle, structural. The check does not distinguish a litigation procedural history from a timeline dump, so the number cannot be driven to zero. The fix belongs in bin/lint-scan.py, excluding litigation/ procedural-history sections. Note that this run added a dated subsection heading to a litigation page — ### Summary judgment and final relief (August 27, 2026) — which is exactly the legitimate case the check cannot tell apart, and the wiki-wide count is unchanged at 28 because the heading is parenthetically dated rather than date-led.
• Carried unchanged. models/deepseek-v3 (in-degree 49 on 608 words, routed to model-robustness-check on 2026-08-22, still unaddressed); the ARIA Charts Code of Practice and the IFPI chart-eligibility principles, both unlocated; TeamT5; the White House release of 2026-07 from the primary-document baseline, instrument still unidentified after nine runs; Amir Salek; the International Catholic Legislators Network; the concept targets parked under needs-review as coined-term and overlap questions.

Dedupe drops (all correct)

Four lane-1 hits already carry open INGEST- tasks and were dropped without re-verification: OpenAI's back-to-school learning report, the Alabama Attorney General's subpoena of OpenAI, Senator Hawley's letter to Flock Safety, and the ITIF/Public First data-center polling deck. That is four of the six lane-1 misses accounted for by work already in flight.

Checks run before closing

• No broken outbound links introduced. Every wikilink on the three expanded pages resolved against the full page set after writing: zero unresolved on all three.
• Style scan clean. All three pages checked programmatically against the banned-header, hyperbole, self-reference, standing-Predictions and date-led-header lists. One apparent hyperbole hit on AI and Civil Liberties — "unprecedented" — sits inside a direct quotation from CDT's Samir Jain and is permitted as quoted material; confirmed to be the only occurrence. Wiki-wide after: banned_headers 0, self_refs 0, predictions 0, header_hype 0, date_headers 28 — none added.
• Revision fidelity. All three edits were additive — new sections appended, frontmatter fields corrected, two infobox rows and one status paragraph rewritten in place. No page was rewritten wholesale, so no _revision-backups/ copy was required. On Anthropic v. United States (Pentagon ban challenge), the pre-August account of the split preliminary-injunction posture was retained in full and re-tensed to past, with a sentence marking where the earlier position ends; no sourced fact, quotation or citation was dropped from any of the three.
• Confidence set honestly. No rating was raised. Anthropic v. United States (Pentagon ban challenge) holds at high, now on primary documents rather than triangulated reporting. Both concept pages hold at medium; neither meets the three-independent-source bar for its core claims, and AI Diffusion's new figures are fast-decay quantitative claims that will need reinforcement within three months.
• Cap respected. Six gaps actioned against a 6–10 cap; 24 candidates found. Five scrapes and two searches, no crawls, no failures.

One-line summary

Three primary documents verified and queued and three high-in-degree pages expanded, on the day a case the wiki has tracked through news reports since March reached final judgment — and the run's more durable finding is that the digest cited the four-page relief order under the label of the 59-page opinion, a mismatch no mechanical check could see, so both are now queued and the opinion's substance is on the page for the first time.