Gap Scan — 2026-08-16

Scanned

Recent window: 5 New Developments Log/ files inside 48 hours (2026-08-13 08:12 and 22:05; 2026-08-14 22:05; 2026-08-15 08:08 and 22:08) and the content pages carrying last_updated: 2026-08-15.

Rotation slice: 3 — concepts/ P–Z, 85 pages, scanned for in-degree-weighted thinness and confidence-versus-sourcing mismatch.

Broken-link analysis: 60 distinct broken targets surviving the escaped-pipe correction. Maximum in-degree was 2; none reached 3. All 28 two-inbound targets went through alias resolution.

Queue state at start of run: empty apart from README.md and the excluded subdirectories — the 2026-08-15 queue-processor pass drained both open tasks.

Run-order note: New Developments Log/2026-08-15-2208-ai-developments.md was unprocessed at the start of this scan; the 2026-08-15 Developments-Log operation covered files through 2026-08-15-0808. Four of the seven gaps actioned below — Inherent, Faraday, SB Energy, and the Faraday arXiv ingest — draw on that unfolded file, which is why they scored as untouched live threads. The nightly fold should still run over it and may find these items already handled here.

Gaps actioned (7 of 31 found)

New pages created (live)

• Inherent — score 5 (live thread +3, core area: frontier-lab governance +2). A London AI lab named in the 2026-08-15 22:08 digest as the developer of a 27B agent reported to beat Claude Opus 4.8 and GPT-5.5, with no page in companies/ and no mention anywhere in the wiki. Alias resolution found nothing under "Inherent," "Faraday," or "Replica." Built from the company's own research post, the arXiv record, tech.eu, Dealroom, The Next Web, UKTN and Cambridge Bio Capital: founding (Tantum Collins, Edward Hughes and Louis Kirsch from Google DeepMind, Kaloyan Aleksiev from Reka AI and Microsoft), the $50M May 2026 seed with a Snapshot funding table, Matt Clifford as adviser, the stated approach, and a Naming section. 1,003 words, confidence: medium, 6 sources. Depends on queued source: arXiv:2608.13331.

• Faraday — score 5 (live thread +3, core area +2). The 27B agent itself. Built from the arXiv abstract record and the developer's research post: the Replica task space (310 tasks, 100 papers), the rubric-based LLM judge with multi-sample aggregation and turn-level credit assignment, the reported cross-lab comparison and its harness settings, and the small-model-directs-large-model result (Faraday calls GPT-5.5 Codex as a tool after training against GPT-5.4-mini). 1,213 words, confidence: medium, 3 sources, open_weights: no, safety_case: none-published. The page records that no numeric scores were published and no independent evaluator has reproduced the comparison. Depends on queued source: arXiv:2608.13331.

• Rosen v. Cook (Apple copyright shareholder-derivative suit) — score 5 (live thread +3, core area +2). The fifth US copyright shareholder-derivative action, folded into AI Copyright on 2026-08-15 as a prose cluster but with no case-tracking page, where the parallel Microsoft action already has Anderson v. Microsoft (copyright shareholder-derivative suit). Built from the as-filed 60-page complaint (12 pages read) with docket metadata confirmed on CourtListener: full party list (fourteen individual defendants plus Apple as nominal defendant), plaintiff's counsel, the three lines of allegation (Books3 via RedPajama; Panda-70M from 3,098,462 YouTube videos; commercial voice models under Illinois BIPA), the eighteen prior actions the complaint pleads as red flags, jurisdiction and venue, and the theory's four prior applications. 1,396 words, confidence: medium, 3 sources. Depends on queued source: the complaint.

• SB Energy — score 5 (live thread +3, core area: compute/infrastructure +2). Named in the 2026-08-15 22:08 digest as the counterparty to Nvidia's discussed $3B equity investment, and already named in prose on OpenAI and Data Center Siting / AI Power Politics since June 2026 without a page. Built from the company's own site and March 2026 partnership announcement plus the two Reuters reports: a Snapshot financing table, the Studio 151 acquisition, the Milam County and Ohio campuses, and the Brockman and Sathe quotations. 916 words, confidence: medium, 4 sources. All August figures are recorded as terms under discussion, not agreements.

Pages expanded (live)

• Surveillance Technology — slice-3 thin anchor, score 4 (in-degree 12 → +2, core area +2). Was 570 words at confidence: medium on sources_count: 1, with its entire "Policy approaches" section carrying unsourced assertions about state face-recognition bans, export controls and constitutional litigation. Added ### Face recognition regulation in the United States (the CDT/Tech Policy Press state survey: fifteen states by end-2024 up from twelve in 2022, the three-tier grouping by name, the five cross-cutting rule counts, Montana's and Utah's first-in-the-nation warrant requirements and how they differ, Maryland SB 182, New Jersey v. Arteaga, the Williams–Detroit settlement terms, and California's 2023 lapse) and ### Commercial spyware and export controls (the November 2021 Entity List addition of NSO Group and Candiru with the Raimondo statement, and the Berkeley CLTC assessment). Added an ## Open questions section, four typed relationships including deployed-by: [[companies/flock-safety]] and regulated-by: [[concepts/export-controls-ai]], and a real Sources list replacing the "Stub created 2026-05-11" line. 570 → 1,613 words; 5 new (Source: URL) cites; sources_count 1 → 6; confidence held at medium. Every pre-existing fact, the FreeForm/Flock paragraph and its 404 Media citation, and the sub-category table survive unchanged. Original backed up to _meta/_revision-backups/surveillance-technology-2026-08-16.md.

• Graph wiring for the four new pages (6 pages touched) — AI Copyright (the Rosen sentence wikilinked and corrected: the suit was filed August 14 and reported August 15, where the page had recorded the report date as the event date; the party count, docket number and the BIPA count added, and the distinction from Anderson's disclosure theory stated); Apple (a full Rosen paragraph in Litigation and regulation, plus the Hendrix, Ted Entertainment and Lopez actions the complaint identifies, and a litigates: relationship); AI for Science (an Inherent subsection under Canonical examples and two new relationships; sources_count 15 → 17); OpenAI and Data Center Siting / AI Power Politics (the bare "SB Energy" mentions wikilinked at first use). last_updated bumped on all six.

Link aliases fixed

None this run. Every multi-inbound broken target resolved to either a genuine but low-priority person/institution gap or a target with no page under any name. No alias errors were found.

Queued — foundational sources

• Falck, Sabri, Surina, Foster, Sims, Devlin, Rogers, Collins, Aleksiev, Kirsch and Hughes, "Training AI Scientists to Replicate Research" — the primary text behind both new Inherent pages, currently cited only as (Source: URL). Verified: arXiv (arxiv.org), the authors' own submission; identifier arXiv:2608.13331 [cs.LG], DOI 10.48550/arXiv.2608.13331, v1 submitted 2026-08-13 14:59:27 UTC, 47 pages. Corroborated against the developer's research post, which links the abstract by ID, and against the May 2026 funding coverage for four of the eleven author names. Raw file: not saved — PDF fetch deferred to ingest to stay inside the daily web-pull cap. Verification record: Wiki/_meta/queue/gap-scan/proposed-sources/training-ai-scientists-replicate-falck-2026.md. Queued: INGEST-training-ai-scientists-replicate-falck-2026.md.

• Verified Stockholder Derivative Complaint, Rosen v. Cook et al., N.D. Cal. No. 5:26-cv-08463 — the primary text behind the new litigation page, with direct precedent in Anderson v. Microsoft — Shareholder Derivative Complaint (June 2026). Verified with a provenance flag: the only reachable full copy is hosted by the copyright-litigation tracker Chat GPT Is Eating the World, since N.D. Cal. serves the as-filed document through PACER. Corroboration was raised accordingly — the CourtListener docket record confirms case name, court, docket number and filing date all four matching the PDF exactly, the PDF carries the ECF stamp "Case 5:26-cv-08463 Document 1 Filed 08/14/26" on every page, and three distinctive passages were confirmed present in the fetched text. Raw file: not saved — 60-page PDF, fetch deferred to ingest; 12 pages read for verification. Verification record: Wiki/_meta/queue/gap-scan/proposed-sources/rosen-v-cook-complaint-2026.md. Queued: INGEST-rosen-v-cook-complaint.md.

Authenticity-verification failures

None. Both sources pulled this run passed all six protocol steps. Three discrepancies were flagged rather than resolved, and none is grounds for rejection:

• The complaint contradicts itself on Cook's announcement date. ¶14 states "April 22, 2026"; ¶22 states "April 20, 2026." Both are recorded on Rosen v. Cook (Apple copyright shareholder-derivative suit) and the conflict is written into the ingest task as a contradiction check against whatever date Apple carries.
• "Faraday" names two different things across sources. The May 2026 funding coverage describes Faraday as the platform Inherent is building; the August 2026 paper uses the name for a specific 27B model. No source reconciles them. Recorded as an open question on both new pages rather than resolved by inference.
• Company-name drift in the digest. The 2026-08-15 22:08 file renders the company "Inherent Laboratories." No primary source uses that form — the company, arXiv, tech.eu, Dealroom, The Next Web and UKTN all render it "Inherent," at the domain inherentlabs.ai. The wiki pages use "Inherent." This is a source-fidelity class-3 residue (an actor rendered other than as the source names it) and is worth a lint check across the digest backlog.

Needs review

None opened this run. The 22 standing notes in queue/gap-scan/needs-review/ are unchanged.

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

• Slice-3 thin anchors not reached — 5 of the 6 the slice surfaced (score 3–4 each). concepts/two-conditions-three-weapons-manipulation (in-degree 11, 988 words, medium/1 — the strongest remaining), concepts/transaction-costs-coase (in-degree 8, 735 words, medium/1), concepts/three-theories-of-victory (in-degree 7, 1,034 words, medium/1), concepts/standards-industrial-complex (in-degree 5, 397 words — the only one under the word bar), concepts/seeing-like-a-state (in-degree 4). All five carry sources_count: 1 and all were last touched 2026-06-06. Also noted but below the thinness rule: concepts/scaling-laws at in-degree 95 and only 806 words, the highest-in-degree page in the slice.

• Live-thread items named in the window with no page (score 3 each). entities/jane-street — the trading firm disclosed a $15B July loss traced in part to its Situational Awareness stake, its first negative trading month since 2016, and is named across the AI-market-exposure thread. models/deepseek-v4-pro and DeepSeek Harness v0.1 — folded into DeepSeek V4 Pro / V4 Flash and DeepSeek on 2026-08-15, which is the right placement until the harness accumulates independent coverage. entities/nevada-transportation-authority — folded into Transportation — AI Deployment; a regulator page would be warranted if it issues further AI permits.

• concepts/replica-benchmark (score 3), deliberately not created. Precedent exists for a benchmark page in concepts/ (CyberGym, Vending-Bench), but Replica has exactly one source — the paper that introduces it — and no independent evaluator has run it. Covered inside Faraday instead, which is where CyberGym's ExploitBench and ExploitGym were also left on 2026-08-15. Revisit when a second party reports Replica scores.

• The escaped-pipe scanner artifact. A naive \[\[([^\]\|#]+) capture reads the escaped pipe in [[folder/slug\|Alias]] links as a trailing backslash and reports ~86 phantom broken targets. The corrected scanner reports 60. This matches the finding recorded on 2026-08-13, -14 and -15; a future run reporting a first-pass figure above ~150 is running an uncorrected scanner.

• Author and institution pages implied by sources/ pages, none created (score 2 each), sixth consecutive run. entities/kevin-klyman, entities/orin-kerr, entities/alondra-nelson, entities/eric-goldman, 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/dawn-song. Better handled as one batch pass than two or three per run; recording it a sixth time without action is itself the finding.

• Institution shorthand targets (score 1–2), unchanged. entities/carnegie-mellon, entities/santa-fe-institute, entities/sais, entities/uw, entities/federal-reserve, entities/ecb, entities/bank-of-england, entities/doj, entities/cma, entities/ec. Several are abbreviations better rewritten to full names than created as pages — sais and uw in particular.

• 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 (fifth consecutive run).

• index.md operational-path backlog (score 1). Twelve of the wiki's broken [[wikilinks]] are index.md pointers to paths that moved under _meta/ in the v4.3 reorganization — Wiki/log, Wiki/lint-report, seven analysis/lint-report-* entries, briefings/2026-05-10-brief, _meta/briefings/weekly-2026-W19, plus one malformed sources/... literal. Index maintenance, not a content gap; carried across five runs.

• Persons named in this window with no page (score 1 each). Damon Falck, Samer Sabri, Anja Surina, Thom Foster, Anya Sims, Sam Devlin, Dylan Rogers (Inherent authors); Tantum Collins, Edward Hughes, Louis Kirsch, Kaloyan Aleksiev (Inherent founders — the strongest of the set, each named on two new pages); Matt Clifford; Abhijeet Sathe (SB Energy); Phil Rosen, Francis A. Bottini Jr. (Rosen counsel); Delsia Bare and Ida Huddleston (the Kentucky landowners). Jake Laperruque and Danny Rimer are each now named once.

One-line summary

Four pages created live — Inherent, Faraday, Rosen v. Cook and SB Energy — one slice-3 thin anchor nearly tripled with real sourcing, six pages rewired, and two foundational sources verified and queued; the user's review is wanted on the two INGEST- tasks and on the digest's rendering of "Inherent Laboratories," a company name no primary source uses.