AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-07

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

Contents

Scanned

Recent window: 4 New Developments Log/ files from the last 48 hours (2026-09-05 22:04 through 2026-09-06 22:09), 56 wiki pages with last_updated inside the window. Rotation slice: 11 — sources/ A–H (374 pages).

bin/lint-scan.py run fresh: 212 distinct broken wikilink targets across 212 references, 0 referenced 2+ times. bin/primary-doc-scan.py --days 3: 6 digests, 12 cited primary-document URLs, 1 missing from Raw Sources/, 1 possible match under another address, 4 lane-2 keyword hits.

Gaps actioned (8 of 27 found)

New pages created (live)

None. The two strongest new-page candidates were declined at the quality gate rather than built thin — see the Deferred backlog. This run's whole budget went to the primary-document lane, which is where the last-48h signal was: the 2026-09-06 22:09 digest is unprocessed at scan time, so its page-level work belongs to developments-log, while the six documents underneath it belong here.

Pages expanded (live)

Both are sources/ A–H pages from the rotation slice that carry substantial bodies but had no ## Relationships section, leaving two of the slice's three highest-in-degree pages with zero typed outbound edges. Prose was not altered on either; no fact or citation was touched.

  • China and the US Are Running Different AI Races — score 3 (in-degree 38, the highest in the slice; US–China competition core +2). 664 words, confidence: medium, sources_count: 1. Added 8 typed links (2 supports, 2 contradicts, depends-on, 2 related), including the two genuine tensions the page states but did not encode: the reported 67-percent Chinese versus 34-percent US industrial deployment rate contradicts AI as Normal Technology's uniformly slow-diffusion account, and the two-races framing contradicts the racing-versus-not-racing axis in Eight Worlds Framework. The existing ## Relevance section, whose six bare unqualified wikilinks are the subject of the standing needs-review note 2026-08-22-ambiguous-bare-wikilinks.md, was left exactly as found; the new section carries folder-qualified targets alongside it rather than replacing it. last_updated 2026-06-06 → 2026-09-07. Confidence held at medium — no new source was folded.

One prose line was changed. The ## Provenance section read "This page summarizes the essay The Adolescence of Technology from darioamodei.com" — a page linking to itself, which resolves to nothing useful and inflated the page's own apparent in-degree. Replaced with the plain title plus a pointer to the raw file it summarizes. No sourced fact or citation was removed.

Zero — eleventh 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; the single-reference long tail was not re-audited, per the settled finding.

Queued — foundational sources

Six, all gap type 9. Four saved to Raw Sources/ in full; two URL-only under the large-document convention. Six verification records written to Wiki/_meta/queue/gap-scan/proposed-sources/.

  • Jakub Pachocki, "An Alien Mind" (OpenAI, 2026-09-06) — score 8 (live +3, primary document +3, safety core +2). The lead item of the unprocessed 2026-09-06 22:09 digest rests entirely on it, and no wiki page names it. Verified: openai.com, first-party, HTTP 200; byline, cover date and both quoted sentences confirmed; cross-linked from OpenAI's own companion post of the same date. Saved in full (including both footnotes): Raw Sources/An Alien Mind (Jakub Pachocki, OpenAI, 2026-09-06).md. Queued: INGEST-pachocki-an-alien-mind-2026.md. Beyond the two sentences coverage quoted, the essay carries a first-party comparative alignment claim for Astra against GPT-5.6 Sol, an account of the Hugging Face agents' behaviour that differs in emphasis from OpenAI's own technical report, and an attribution of "recent cybersecurity incidents involving a non-OpenAI model" to motivated reasoning under optimization pressure — a characterization of another developer's incident that the ingest task flags for attribution rather than statement.
  • OpenAI, "Research acceleration: The view inside OpenAI" (2026-09-06) — score 8. Same digest, second item. This would be the first quantified first-party disclosure the vault holds of how far a frontier lab's own research is automated, and OpenAI argues it should be made mandatory, which places it past a press release. Verified: openai.com, HTTP 200; all six figures the digest attributes to it read off the document. Saved in full for text and the methods appendix: Raw Sources/Research acceleration - The view inside OpenAI (OpenAI, 2026-09-06).md. Queued: INGEST-openai-research-acceleration-2026.md. Two things the coverage does not carry: the substitution finding — restricting Astra-class compute left total RL allocation largely unchanged because other model classes absorbed about 85 percent of the decline — and OpenAI's own methodological caveats, including that experiment-velocity growth is confounded by compute growth since 2025. One internal discrepancy recorded rather than resolved: the prose says 3.1 agent-workdays per human workday, the chart annotation reads 3.14x.
  • OpenAI, "PRC-linked influence operations are targeting AI debates in the US" — score 8 (live +3, primary document +3, national-security core +2). Named under ## Primary documents in the 2026-09-06 08:09 digest, where Wired's account of the data-center backlash rests on it. Verified: openai.com, HTTP 200, complete; campaign names "Data Center Bandwagon" and "Tech and Tariffs" and the prompt instruction excluding Xi Jinping confirmed verbatim. Saved in full: Raw Sources/PRC-linked influence operations are targeting AI debates in the US (OpenAI, 2026-06-10).md. Queued: INGEST-openai-prc-influence-operations-2026.md. The document is dated June 10, 2026 — see Source-fidelity flags. The underlying full threat report is a separate CDN PDF, not pulled.
  • FCA, "Frontier AI and cyber resilience" multi-firm review — score 6 (live +3, primary document +3). The evening digest's International AI Regulation item is about it. Verified: fca.org.uk, the regulator's own domain, HTTP 200; the finding the digest attributes to it is insight theme 1 verbatim. Saved in full: Raw Sources/Frontier AI and cyber resilience - multi-firm review (FCA, 2026-08-26, updated 2026-09-02).md. Queued: INGEST-fca-frontier-ai-cyber-resilience-2026.md. Published August 26, not September 2 — see Source-fidelity flags. The review states in terms that it "does not introduce new rules, guidance or regulatory expectations" and summarises what firms reported rather than the regulator's own testing; both limits are carried into the ingest task, since a fold that presented it as an FCA requirement would reverse its modality.
  • "A Definition of AGI" (Hendrycks, Song, Szegedy et al., arXiv:2510.18212) — score 8 (live +3, primary document +3, frontier/safety core +2). Gary Marcus pointed Jensen Huang to this document as the definition Huang's AGI declaration lacked. The vault holds four concepts/ pages in this area and no copy of the work; a search for its title and its arXiv identifier across all content pages returns zero hits. Verified: agidefinition.ai, the authors' own project site under the Center for AI Safety mark, HTTP 200 — full 33-author list with 30 numbered affiliations, the CHC-grounded ten-component framework, a per-component results table, and a BibTeX block giving the identifier. URL-only; fetch the arXiv PDF at ingest. Queued: INGEST-hendrycks-definition-of-agi-2025.md. It is a 2025 paper, not a September 2026 statement — see Source-fidelity flags.
  • Silbey and Hartzog, "AI Slop" (SSRN 5870742) — score 6. Named under ## Primary documents in the morning digest, which carried the revision as its own item. AI Content Saturation (\"AI Slop\") exists and was edited inside the window; the vault holds no legal-academic treatment of the concept and not the paper that proposes the definition the digest quotes. Verified: SSRN, the standard host for an unpublished law-review preprint, HTTP 200; title, both authors, both BU affiliations, 70-page length, all three dates, and the quoted definition and its three constituent parts confirmed from the abstract page. No DOI is issued, which is expected for an SSRN working paper and is not a failure. URL-only. Queued: INGEST-silbey-hartzog-ai-slop-2025.md.

Source-fidelity flags

Three of the six documents are misdated in the digests that name them, all in the same direction — the date of the coverage standing in for the date of the document. This is the send-date-as-event-date class named in source-fidelity v4.7, but arriving one layer further out: through a digest's ## Primary documents list, which the v4.15 capture rule added and which records URLs without dates. The list is doing its job — all three were found because it named them — but a bare URL carries no date, so the date is inherited from the item that cites it.

  • OpenAI PRC-linked influence operations — listed undated, cited by Wired on 2026-09-06, in a context reading as a contemporaneous disclosure. Cover date June 10, 2026, roughly three months earlier.
  • FCA multi-firm review — the digest says "published a multi-firm review on September 2, 2026." Page metadata gives article:published_time 2026-08-26 and funnelback:created-date 26/08/2026, with September 2 as article:modified_time. August 26 is publication; September 2 is a revision.
  • A Definition of AGI — referred to as "the agidefinition.ai statement," which reads as current. The authors' own BibTeX gives year 2025, and the 2510 arXiv identifier places submission in October 2025. A second flag on the same document: "statement" is the digest's word. It is a research paper with an author list, affiliations, a framework and results, so the 33 names are co-authors, not signatories, and its source_type is paper.

None of these is an error in the digests' prose, and all three are recoverable from the primary document — which is the argument for capturing rather than folding. Whether the ## Primary documents section should carry a date field alongside its URLs is a New Developments Log/SCHEMA.md question and is not actioned here, since a schema change ships alone.

Authenticity-verification failures

None. Six documents sought, six verified, all on a first-party or canonical host on the first attempt — three on openai.com, one on fca.org.uk, one on the authors' own project site with arXiv as the ingest target, one on SSRN under an author-granted license. No mirror, aggregator or news reproduction was used, and no provenance flag was needed.

One document was declined rather than failed: the X Global Affairs post naming roughly 200 China-linked anti-data-center accounts (x.com/globalaffairs/status/2093130747796148634), named beside the OpenAI threat post in the same digest. A platform post behind an authentication wall is not reliably capturable and its provenance cannot be established to the protocol's standard from outside. Recorded in the raw file and the ingest task as a curator decision.

Dedupe drops (all correct)

  • US statement of interest at gov.uscourts.nysd.640396.1682.0.pdf — the only lane-1 hit, and the third consecutive run on which it is reported. It is the second docket copy of the statement of interest already held as Raw Sources/Statement of Interest of the United States - In re OpenAI Copyright Infringement Litigation (DOJ, 2026-09-01).md, captured from the 612697 address; the raw file's own text carries both docket numbers. Not a gap, and it will keep surfacing: a document docketed in two MDL member cases has two URLs and one raw file, which primary-doc-scan.py matches by address. Worth a one-line URL-alias list in the scanner rather than a fourth manual drop.
  • METR and Redwood Research, Hugging Face incident report (August 2026) — the scan's "possible match under another address" hit. Confirmed held as Raw Sources/Brief Independent Investigation of Agents' Behavior in the OpenAI - Hugging Face Incident (METR and Redwood Research, 2026-08-26).md. The digest cited metr.org/hugging-face-incident-report-aug-2026.pdf; the vault holds it under its title. This is the exact shape the v4.15 changelog predicted the third state would catch, and it worked.
  • OpenAI, Hugging Face incident technical report — held as Raw Sources/OpenAI - Hugging Face Incident Technical Report (OpenAI, 2026-08-26).md.
  • OpenAI, Preparedness Framework v2 — held as Raw Sources/OpenAI Preparedness Framework V2.md.
  • GPT-6 Astra deployment-safety documentation — the deploymentsafety.openai.com PDF is the system card, held in full and carrying an open INGEST-gpt-6-astra-system-card-2026.md since 2026-09-06.
  • SCO Bishkek Declaration — the digest's own primary-documents entry records "item not carried," so there is no wiki claim resting on it.

Four lane-2 keyword hits were read before dropping, per the instruction that lane 2 is a pre-filter: the Seattle Times/Newsday complaint (held, and its litigation and source pages created 2026-09-05), the White House polysilicon proclamation (the Wacker Chemie item quotes it but is not about it), the GPT-6 Astra system card (the evaluation-revision item is about the announcement page and its archive snapshots), and the ChinaTalk open-weights essay (an argument, not an instrument).

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

  • The sources_count: 1 thin-anchor diagnostic does not work on sources/ slices. The standard scan flags 134 of 374 pages in this slice. Almost all of that is structural: a source summary page summarizes one source, so sources_count: 1 is correct by design, and flagging it produces a list that is 90-percent noise. This is the third sources/ slice in the rotation (11, 12, 13) and the diagnostic will misfire on all three. What the scan does legitimately surface in these slices is the confidence: high on sources_count: 1 pattern — 20 of the 25 highest-in-degree flagged pages here, led by Statement on AI Risk (CAIS), Alignment Faking in Large Language Models and FLI — Pause Giant AI Experiments: An Open Letter, none reinforced since 2026-06-06. This is the fifth consecutive report of that pattern, now in a seventh folder. The 2026-09-06 recommendation stands: a single confidence-audit pass would close more of it than one page a day, and this skill's cap structurally cannot get there.
  • FCA / Bank of England / HM Treasury joint statement on frontier AI models and cyber resilience (May 2026) — score 5, new this run. Discovered inside the FCA review, which builds on it ("In May 2026, we said these models represent a step-change in capability"). A tri-authority statement is the more citable instrument of the two and the vault holds neither. First candidate for the next run's queue. Also surfaced and not held: CMORG "Firm Guidance For Frontier AI" (June 2026), NCSC frontier-AI guidance, the G7 cyber expert group statement (September 2025), the 2025 CBEST Thematic.
  • OpenAI June 2026 Threat Report PDF (score 4), new this run. The full document behind the PRC influence-operations summary captured today. Deferred rather than pulled: the summary carries the claims the wiki needs now, and the report is better fetched whole at ingest.
  • Two chain-of-thought monitoring papers, arXiv:2507.11473 and arXiv:2503.11926 (score 4), new this run. Both cited by Pachocki as the technical basis for OpenAI's primary monitoring bet, whose diminishing reliability is the essay's substantive disclosure. Neither is held. The strongest of the nine dangling references the essay creates.
  • Epoch AI AI-R&D task taxonomy (score 3), new this run. The analytical frame for section 3 of the research-acceleration disclosure; not held.
  • Supreme Court of India, Vijay Ghanshyam Gadiya v. Union of India (score 5), new this run. Named in the evening digest's primary-documents list, cited by the Ctrl+AI+Reg roundup. Not pulled: the sci.gov.in judgment URL is a query-parameterised viewer, the item was not carried in the digest's prose, and the vault has no Indian AI litigation coverage to land it in. A curator decision on scope more than a capture task.
  • ARC Prize Astra evaluation post (score 3), new this run. Cited by Exponential View for ARC-AGI-3 action-efficiency figures the wiki carries secondhand on Astra. The vault holds the ARC-AGI-3 challenge announcement but not the Astra evaluation. Naturally belongs with the queued Astra system-card ingest.
  • Swiss Re data-center insurance report (score 2), new this run. The evening digest records "no URL given in the reporting; not located." Not a capture failure — there is nothing to attempt until a URL surfaces.
  • entities/keith-ellison (score 3) — second deferral. Named on at least three content pages and the defendant of record in a tracked case. Still no first-party material pulled.
  • models/grok-imagine (score 3) — second deferral. Needs capability and release material this run did not gather.
  • CCW autonomous-weapons consensus document, Geneva, September 5, 2026 (score 5) — second deferral. Held at the quality gate for the same reason as last run: the only available account is one wire report plus one advocacy characterization, and a page built from those would inherit that framing. The agreed text is the right first capture and has not been located.
  • analysis/missing-coverage-2026-04-14 and analysis/crystallize-session-2026-04-13 — second deferral. April 2026 session artifacts sitting in mainspace. A placement decision (analysis/ versus _meta/), not a content gap.
  • models/deepseek-v3 — removed from this backlog per the 2026-09-05 and 2026-09-06 reports. It passes every mechanical thin-anchor test; its problem is depth against importance, which model-robustness-check measures and this skill does not. Not re-listed.
  • entities/adam-thierer (score 3) — tenth consecutive deferral, structural.
  • Date-stamped headers in lint-report.md (score 1) — twenty-second cycle, structural. Holding at 26. Six self-references, unchanged.
  • Carried unchanged. entities/nightingale; legislation/caadca; entities/david-feith and entities/hudson-institute; Marvell (no company page, since 2026-08-21); California AB 1856; the Texas OAG / Meta Webster–Mahoney exchange; the draft executive order on a frontier-developer self-regulatory organization; the AI Evaluator Forum / AEF-1 standard; ExploitBench flag-level disclosure; the "Carolina Principles"; OpenAI's "Our agreement with the Department of War"; the House E&C markup memo; the New York City Public Schools generative-AI policy; the Agentic Commerce Protocol / Model Context Protocol placement question; OpenAI's August 31 response to Rep. Casar (Google Drive host, curator decision); the July 2026 statement from more than 1,000 AI-firm employees.

Run integrity

  • No broken outbound links introduced. Both touched content pages and all six new queue files were resolved against the full 1,800-plus page set after writing: zero unresolved. Eight hedged link targets caught in that check and fixed before finishing — they were written into the ingest tasks with qualifiers like "check the exact slug" and "if it exists," which the resolver correctly reads as links regardless. All eight were repointed to real pages: chain-of-thought-monitoring → Post-Deployment AI System Monitoring, third-party-auditing → Alignment Auditing, frontier-safety-frameworks → AI Safety Cases and Frameworks, international-ai-governance and uk-ai-governance → Frontier AI Governance, influence-operations → AI and Misinformation, ai-generated-disinformation → AI Content Provenance, operational-resilience → Financial Stability Board (FSB), ai-and-authorship → AI Content Licensing, ai-in-scientific-publishing → AI for Science. A hedge is not a licence to write a link that does not resolve.
  • Style scan clean. Both touched content pages checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each.
  • Revision fidelity verified. Both edits were additive apart from one last_updated value each and the single Provenance line described above. No sourced fact, figure or citation was removed from either page; the six bare wikilinks in the china-us-different-ai-races Relevance section survive untouched.
  • Confidence not inflated. Neither expanded page's confidence was raised. Both remain medium on sources_count: 1, which is correct — graph wiring is not new evidence, and the autonomy rules permit a rise only when supporting sources are folded.
  • Capture completeness stated, not assumed. Four of six documents saved in full. The research-acceleration post is complete for text and the methods appendix but its eleven charts render their series client-side; only axis labels and inline annotations were capturable, and the raw file says so rather than presenting itself as a full capture. Two are URL-only under the sanctioned large-document convention (70-page law review article; arXiv paper).
  • Cap respected. Eight gaps actioned against a 6–10 cap; 27 found after dedupe. Six scrapes, no searches, no crawls, no failures.
  • A dedupe drop recurring for the third run is reported as a scanner issue, not re-dropped silently. See the 640396 statement-of-interest note above.

One-line summary

Six primary documents behind the last 48 hours of digests captured and queued — four saved in full — with three date errors caught in the process, all inherited from the coverage rather than the source; two high-in-degree sources/ pages given typed relationship sections; no new pages created, two candidates held at the quality gate; the evening digest is unprocessed for the sixth consecutive run, and the confidence: high on single-source pattern now appears in a seventh folder.

Categories: Briefings | briefing | gap-scan