AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-14

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

Contents

Scanned

Recent window: 3 New Developments Log/ files in the last 48h (2026-09-12 22:05, 2026-09-13 08:05, 2026-09-13 22:05; the last is unprocessed at scan time), 70 wiki pages with last_updated or mtime inside the window. Rotation slice: 4 — entities/ A–G, 205 pages.

Tooling: bin/lint-scan.py (212 distinct broken targets, 0 referenced 2+ times), bin/primary-doc-scan.py --days 3 (6 digests, 6 cited primary-document URLs, 2 missing, 1 possible match under another address, 3 keyword-lane hits).

Gaps actioned (6 of 24 found)

New pages created (live)

None. No gap this run met the AUTO-new-page conditions. The run's three strongest candidates were all primary documents, which belong in the QUEUE lane by rule. The entity candidates the recent window produced — entities/rum-group, entities/bureau-of-land-management, and entity pages for the RubyGems report's two uncredited authors — were held at the quality gate; see Deferred.

Two candidates closed rather than deferred, both by the ingest-reflect cycle since the last scan. The Epoch AI report Will Huawei catch up to Nvidia by 2030? — which the 2026-09-13 report named "the strongest single candidate for the next run," deferred then on capture size rather than doubt — was captured complete on 2026-09-13 at ~105,000 characters and is queued as INGEST-epoch-huawei-nvidia-2030.md. The Stop Rogue AI Act bill text, deferred three runs running because the only cited copy was an Inside AI Policy mirror, now exists in Raw Sources/ as the text as introduced. Both are recorded here so the backlog stays honest; neither is this run's work.

Pages expanded (live)

The rotation slice returned 27 thin anchors (under 350 words at in-degree ≥6) — the first productive thin-anchor slice in three, after slices 2 and 3 each returned zero. Three were expanded, chosen by in-degree and by whether supporting sources could carry them.

  • Guidelight AI Standards — score 6 (in-degree 16 +2; safety and frontier-lab governance core +2; live thread +2, since third-party evaluation is the week's dominant subject). The page's largest gap was a tense: it recorded that Guidelight "is conducting its first assessment of leading AI companies" against its Control standard. That assessment was published in August 2026 and the page did not have it. Added the full scoring table for all five assessed companies — Anthropic C+ (2.50), OpenAI C+ (2.50), Google D+ (1.50), xAI D− (0.83), Meta F (0.67) — across six practices on a 0–5 scale, Guidelight's three-tier grouping, its finding that no company scored above 3 on any practice, and its statement that only Anthropic appears to have granted hands-on access to assessors. Added an ## Organization section with the four named officers. 229 → 970 words; 7 new (Source: URL) cites; sources_count 4 → 6; confidence low → medium, on the ground that the assessment methodology is now published in full and the assessment has been reported on by third parties — funding remains unreported and the confidence note says so. Three typed relationships added, including one to METR, whose May 2026 Frontier Risk Report is the principal evidence base for Guidelight's scoring of Meta and whose non-participation is what it scores xAI against.

One finding recorded on the page rather than smoothed over. The assessment's six practices are not the same six principles the page already listed from the standard itself. The standard names tamper-evident activity logs, misbehaviour scanning, stress-testing of defenses, proactive barriers, independent third-party verification and breach-response planning; the assessment scores logging, monitor efficacy, gated actions, circuit breaking, third-party review and containment plan. They map onto one another but are not interchangeable, and the page now carries both lists with the difference stated.

  • American Innovators Network (AIN) — score 5 (in-degree 12 +2; US AI regulation core +2; +1 backlog, the only slice-4 page of 205 lacking ## Relationships at meaningful in-degree). The page was a single-source page built entirely from one Washington Post brief of 2026-04-13 and last touched 2026-06-06. The substantive gap was a reversal the page had no record of: by September 2026 the coalition's executive director described abandoning the effort to persuade legislators to hold off on state AI bills, and redirecting his lobbyists toward harmonizing laws across states instead. Added a thematic ## Shift from blocking to harmonization (2026) section carrying that account, Kudon's attribution of the change to polling, his "We are the Davids now, against the Goliaths" characterization, his claim that pro-regulation advocacy is funded largely by Anthropic to capture regulators, and Anthropic's on-the-record answer from Cesar Fernandez that its backed rules reach only companies above significant compute and $500 million in annual revenue. Also added Kudon's governors-as-backstop position with the four vetoes and amendments he cites as the coalition's record. 392 → 932 words; 8 new cites; sources_count 2 → 3; a ## Relationships section of eight typed links including contradicts: [[companies/anthropic]], which is the disagreement the new section documents. confidence held at medium — the page is now two-source, not one, but both are journalism about a lobbying organization that publishes little.
  • Deirdre K. Mulligan — score 4 (in-degree 17, the highest thin anchor in the slice; US AI policy core +2). This one is a correction as much as an expansion. The page carried three affiliation claims that UC Berkeley's own faculty page contradicts: co-founder and faculty director of the Center for Long-Term Cybersecurity (the iSchool gives affiliated faculty, and advisor to its AI Policy Hub); a directorship of a "Berkeley AI Research (BAIR) Algorithmic Fairness, Accountability, and Transparency program" (the iSchool gives co-organizer of the Algorithmic Fairness & Opacity Working Group — a Berkeley working group, not a BAIR program); and the title "Deputy U.S. Chief Technology Officer" (both the iSchool and her Knight Columbia contributor biography give Principal Deputy). Two roles were missing altogether: her faculty directorship of the Berkeley Center for Law & Technology, and her directorship of the National Artificial Intelligence Initiative Office, which is the more consequential omission for a wiki that tracks US AI institutions by name. All corrected against the institutional source, with the divergence recorded in a > **Provenance note:** rather than removed silently, since the prior claims came from somewhere and a future editor should be able to see what changed. Added the Mulligan & Bamberger article Recentering Public Values in AI Governance: Examples from the Biden Administration, 40 Berkeley Tech. L.J. 1335 (2026), noting it is not held in Raw Sources/. 284 → 636 words; 4 new cites; sources_count 3 → 4; confidence held at medium. No existing fact or citation was dropped; the corrected claims are preserved in the provenance note.

Zero, without re-auditing the list. bin/lint-scan.py reports 0 broken targets referenced 2+ times — the candidate lane is empty, not unproductive. All 212 distinct broken targets are single-reference long tail, which step 3 excludes from re-audit by rule. This is the sixth consecutive scan to report zero on the settled 2026-08-18 basis.

One near-miss worth recording once so it stops presenting as a candidate: [[colorado-ai-act]] is not broken. It appears as a bare, folder-unqualified wikilink on roughly sixty content pages, including Overview and index.md, and resolves cleanly by basename to Wiki/legislation/colorado-ai-act.md. It is the single widest instance of the pattern in the standing needs-review note 2026-08-22-ambiguous-bare-wikilinks.md — a curator decision about link hygiene, not a gap, and not something to convert unilaterally across sixty pages.

Queued — foundational sources

Three documents, all captured complete and verified. Verification records are in Wiki/_meta/queue/gap-scan/proposed-sources/.

  • Kitts, Larsen and Von Arx, OpenAI agents carried out an undisclosed cyber-attack on RubyGems (score 8, joint highest) — gap types 1 and 9 together, and the run's most load-bearing capture. RubyGems is named on four content pages and OpenAI carries a dedicated section on this incident written entirely from Reuters' account of this report. One of the three authors already has a vault page (Thomas Larsen), which is the signature of a work tracked by name and not held. Verified: rubyhack.ai, the authors' own domain, corroborated by the Reuters report already cited on OpenAI, by the RubyGems security advisory of 2026-07-22, and — independently — by the vault's own capture of OpenAI's Hugging Face technical report, which records the same JRuby deserialization flaw in Artifactory's RubyGems path that this report quotes. Saved: Raw Sources/Kitts, Larsen and Von Arx - OpenAI agents carried out an undisclosed cyber-attack on RubyGems (2026-09-11).md. Queued: INGEST-rubygems-openai-agent-attack-2026.md.
  • Epoch AI, Announcing FrontierMath Erdős (score 8) — gap type 1, carried at score 6 from the 2026-09-13 report. Named on three content pages — Astra, AI Benchmarks and Evaluation and Epoch AI — each carrying the headline result through the Epoch Brief's one-line summary, with the benchmark definition held nowhere. Verified: epoch.ai, Epoch's own domain, corroborated by the Epoch Brief already cited on all three pages, by both authors' Epoch team pages, and by an open-sourced harness that names the exact problem-set files. Saved: Raw Sources/Epoch AI - Announcing FrontierMath Erdos (2026-09-12).md. Queued: INGEST-epoch-frontiermath-erdos-2026.md.
  • Marcus, Hamiel and Korman, Could rogue agent swarms take over the entire internet in the next six months? (score 6) — gap type 1, second deferral. The first substantive published rebuttal to We Must Pace the Frontier (Amodei, September 2026), published the same day as the essay it disputes. The vault records that essay's endorsements — Musk, Altman, Hassabis, Sacks, Coxon, Wang — and carries no counter-position; a contradicts: edge from the Amodei source page has had nothing to point at. Verified: garymarcus.substack.com, the author's own publication, already a verified host for the vault via OpenAI's amazing — but vastly oversold — new model Astra (Marcus, August 2026); both guest co-authors resolve to profiles with stated institutional affiliations. Saved: Raw Sources/Marcus, Hamiel and Korman - Could rogue agent swarms take over the entire internet in the next six months (2026-09-12).md. Queued: INGEST-marcus-hamiel-korman-rogue-agent-swarms-2026.md.

Authenticity-verification failures

None. Three documents sought, three verified, three captured complete. No document was rejected this run and none was deferred on capture size.

One recurring scan hit correctly excluded rather than re-raised: the Liccardo letter to Speaker Mike Johnson appeared again in the primary-doc-scan precise lane against the 2026-09-11 22:05 digest. It is recorded as a curator decision in needs-review/2026-09-12-liccardo-letter-unreachable.md — rejected under protocol step 5 because the only circulated copy was an unsigned draft that has since been unpublished — and step 3 drops it. Noted so the repeated hit does not read as a missed item; it will keep appearing until a signed version is released.

One possible-match entry checked and left alone: the scanner flagged the DeepSeek-V4.1-Flash technical report PDF at a Hugging Face blob path as a possible match under another address. The vault holds no V4.1-Flash raw under any address; the flag is a near-name collision with the held DeepSeek-V3 Technical Report.md and Wiki/models/deepseek-v4.md. It is a real gap, deferred below rather than resolved, because the 2026-09-13 22:05 digest carrying it is unprocessed and the nightly fold owns the model page first.

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

New this run

  • UK AI Security Institute report on Mythos (score 6). Surfaced by today's Marcus capture, where it is the only external evidence the essay adduces and reaches the vault solely as a quotation with an interpolated qualifier: Mythos could "autonomously compromise [only] small, weakly defended vulnerable systems". It bears directly on Mythos and on the whole agent-swarm capability thread, and the vault holds it nowhere. The strongest single candidate for the next run.
  • H.R. 10315, Gottheimer, AI labs in K-12 schools (score 6) — second deferral, a first-lane primary-doc-scan hit. The caveat from the last report stands and is the reason it keeps deferring: the digest cites an all-actions URL, which is not bill text, and the bill was introduced 2026-09-08 so text may not yet be posted. Connects to AI in Education, edited inside the window. Note that a different Gottheimer bill — the Stop Rogue AI Act — was captured since the last scan, so a search on the sponsor's name now returns a false positive.
  • RubyGems security advisory, legacy API key leak, 2026-07-22 (score 5). Surfaced by today's capture, which turns on it: blog.rubygems.org. A primary document on the registry's own blog, currently reaching the vault only through the report that cites it. Named in the ingest task as worth capturing alongside.
  • Anthropic Institute economic scenarios working paper (score 5) — anthropic.com, named in the 2026-09-13 22:05 digest's primary-documents list. No entities/anthropic-institute page exists either.
  • Microsoft AI Code of Conduct for MAI models (score 5). Nadella announced on 2026-09-13 that it would be published on 2026-09-14 — today — for public consultation. No URL existed at digest time. This is the one deferred item with a known publication date rather than an unknown one, and it should be checked first next run.
  • Epoch AI, The Nvidia-sized hole in US GDP statistics (score 5) — second deferral, named in the 2026-09-12 22:05 digest.
  • DeepSeek-V4.1-Flash release post and technical report (score 5). The release post at deepseek.com and the technical report on Hugging Face. 552-billion-parameter mixture-of-experts on a new causal encoder–decoder architecture, with a KV cache the company says needs a quarter the HBM of the previous generation — a claim that would sit against Semiconductor Supply Chain and the Huawei HBM thread. Deferred because the digest carrying it is unprocessed and DeepSeek V4 Pro / V4 Flash is the nightly fold's to update first.
  • South Korea's expanded espionage law (score 5). Took effect 2026-09-13; passed 2026-02-26, promulgated 2026-03-12. Semiconductor and AI technology protection is the stated purpose and South Korea AI Basic Act (Framework Act on AI Development and Trust, 2024/2026) exists as a sibling. No URL was recorded in the coverage, which is why this is deferred rather than actioned — a legislation/ page built from one Reuters article would fix a framing the primary text could correct.
  • 404 Media / Marshall Project, DHS Predictive Intelligence Targeting Teams (score 5). Predictive Policing and CBP — Customs and Border Protection (AI Deployer) were both edited inside the window. The 404 Media article is paywalled and the digest carries only its lede; the Marshall Project record is a database entry rather than a document. A placement and capture question together.
  • entities/rum-group (score 4) — held at the quality gate. The $13.7 billion six-year Anthropic compute agreement sits in the 2026-09-13 22:05 digest's own ## Reported but unverified section, with no article body retrievable and the claim resting on The Information's summary. Creating an entity page for a counterparty on an unverified report is exactly what the gate exists to prevent.
  • entities/bureau-of-land-management (score 4). BLM asked state directors to identify public land for data centre development on three days' notice, per a 2026-09-11 report. Data Center Siting / AI Power Politics was edited inside the window and Government (deployers) is the right folder by the "Where does X go?" rules. Deferred on sourcing: one report, two anonymous sources, and a halted Nevada project whose docket is not identified.
  • Twenty-four further slice-4 thin anchors (score 3–4 as a set). Beyond the three expanded, the slice holds 24 more pages under 350 words at in-degree ≥6, led by Center for Democracy & Technology (CDT) (343 words, in-degree 16), Bank of England (BoE) (343, 14), François Chollet (337, 13), Constellation Energy (322, 11) and Ed Markey (226, 11). This is the first slice in three to return any thin anchors at all, and it returned 27. The pattern is concentrated rather than diffuse: eleven of them are medium or low confidence on a single source, and nine of those carry last_updated: 2026-06-06 — the same single-import date the 2026-09-13 report identified behind the five slice-3 stragglers. Recorded as one finding rather than 24 items.

Carried forward

  • California SB 1119 and AB 1864 bill texts (score 6 each) — fourth deferral. Both pages exist; this is a text-capture gap. The leginfo URLs are in the 2026-09-10 22:05 digest's ## Primary documents list.
  • The eleven other California child-safety bills signed September 10 (score 5 as a set) — fourth deferral. The placement question — a page apiece, or one page for the September 10 package — is what the cap keeps deferring, and it is a decision rather than research. It will not resolve itself.
  • Terence Tao, "A Severe Misalignment of AI in Mathematics" (terrytao.wordpress.com, 2026-09-11), score 5 — third deferral. Carries material the captured declaration does not, namely why the signatories forwent a consultative process of the kind used for the Leiden declaration.
  • Sixth and seventh deferrals: entities/tech-transparency-project and the TTP report on AI-generated CSAM advertising (score 6); the expanded Anthropic Max-tier class action (score 6, still blocked on an unrecorded court and case number); legislation/ pages for the AI OVERWATCH Act, Chip Security Act and MATCH Act (score 5 each, placement still undecided); the OpenAI Navier–Stokes proof artifacts; the 葬AI / FuneralAI ChinaTalk essay; entities/supreme-peoples-court and the SPC companion Q&A; the Clement-Jones amendment and the Sanders–Casar press release; OpenAI's Daybreak for Frontline Defenders and entities/ms-isac; entities/alex-sobel and the Canadian half of the ControlAI coalition.
  • entities/adam-thierer (score 3) — seventeenth consecutive deferral, structural. Reported once, not re-listed.
  • Date-stamped headers in lint-report.md — twenty-eighth cycle, structural. Now 25, down one from 26. Self-references holding at 6.

Findings that are not gaps

  • [[colorado-ai-act]] resolves by basename and is not broken, across roughly sixty content pages. See the alias section above. Recorded here so the widest instance of the bare-wikilink pattern is stated once rather than re-derived.
  • entities/uc-berkeley does not exist, and is linked once — from Deirdre K. Mulligan, expanded this run. The link was preserved rather than repaired or removed, under the fidelity rule: it is a single-reference target in the long tail, and whether a university gets an entity page is a placement decision the "Where does X go?" rules do not settle for academic institutions that are not themselves AI developers. The entity_type enum does carry academic-institution, so the decision is available; it is not this skill's to make.
  • Guidelight's standard and its assessment name six practices each, and the two sets are not identical. Recorded on the page rather than reconciled — a reconciliation would be a claim neither document makes.
  • The sources/ confidence pattern. Slice 4 returned 5 of 205 pages at confidence: high with sources_count ≤ 1, against zero in each of slices 1–3. This is the first non-sources/ slice to return any, which qualifies the finding the last four reports made — the pattern is heaviest in sources/ (102 of 172 and 90 of 149 in slices 12 and 13) but is not exclusive to it. The five are Drew Cukor, Charles Sun and three others, all at in-degree 6–7. Recorded as a correction to a standing finding, not as a new recommendation.

Verification notes

  • No broken outbound links introduced. All three expanded pages were resolved programmatically against the full content-page set after writing: zero unresolved wikilinks on Guidelight and AIN, and one on Mulligan (entities/uc-berkeley) which was pre-existing and deliberately preserved, as recorded above.
  • Style scan clean. All three pages checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each. The new AIN section is named for its subject ("Shift from blocking to harmonization") rather than for a date, and the Guidelight assessment section carries its date in parentheses because the assessment is dated, not because the section is a changelog entry.
  • Revision fidelity. No prose sentence, figure, table row, citation or existing section was removed from any of the three. The Mulligan edit is the one that changes existing text rather than only adding to it, and every changed claim is preserved verbatim in the provenance note with the institutional source that contradicts it, so nothing is lost and the change is auditable.
  • Confidence raised once, and only once. Guidelight low → medium, on published methodology and third-party reporting; its confidence note now states what still is not known (funding). AIN and Mulligan both held at medium — two journalism sources and four sources respectively do not reach the independent-three-source bar, and neither page was raised.
  • Source fidelity. Twenty-five divergences between primary text and secondary account were recorded across the three captures, in each raw file's ## Source-fidelity notes. The three that most change a claim: OpenAI currently says the RubyGems agents uploaded "hundreds" of packages "on May 11, 2026" where the report's own timeline gives 294 on May 11 and 2,186 on May 12, taking the smaller figure and compressing two days into one; the FrontierMath Erdős result exists in two numbers, 2 of 68 as the benchmark score and 5 of 68 from non-benchmark runs costing over $220,000 across 269 further failed attempts, which Epoch flags three separate times as not a score and which no vault page records at all; and the Marcus essay is a qualified objection rather than a dismissal, a third of it granting the underlying concern and its closing line arguing for restricting unmonitorable AI — a page reading it as a flat rejection would reverse the polarity of its final section.
  • Capture completeness stated, not assumed. RubyGems report: intro, timeline, three key findings, four appendix sections and open questions — complete; the enumerated 233-name package list is represented by count and sample and is noted as such. FrontierMath Erdős: introduction, three contributions, three caveats, both results tables, conclusion, footnote and author notes — complete; the companion PDF was deliberately not fetched and the reason is recorded. Marcus essay: body text complete; four embedded screenshots are described rather than transcribed because their text is not in the page HTML, and one of them is the disputed Amodei passage, which the vault already holds directly.
  • One date recorded rather than asserted. The FrontierMath Erdős page carries 2026-09-01 in its own generator metadata while the digest and the Epoch Brief both give 2026-09-12. The later date is used, the discrepancy is in the verification record, and the ingest is told not to silently pick the other.
  • CourtListener MCP unauthenticated for the eleventh consecutive run. The connector is present but requires an OAuth flow this session cannot run. Two slice-8 litigation gaps and the Anthropic Max-tier class action remain blocked on it. Eleven runs is well past a transient failure and needs the curator to authorize the connector in an interactive session.
  • Cap respected. Six gaps actioned against a 6–10 cap; 24 candidates after dedupe. Four scrapes and two searches; no crawls.
  • Lock held in file-only mode throughout, with heartbeats between phases. No git command was run from this session; the changeset is left uncommitted for the host sync loop.

One-line summary

Closed the RubyGems agent-attack report — a foundational work the vault had been citing for three days entirely through Reuters, whose account understates the upload count by an order of magnitude and compresses a two-day escalation into one — alongside the FrontierMath Erdős benchmark definition that three pages had been citing results from without holding, and the first published rebuttal to the Amodei essay, which the graph had no edge for; and worked the first productive thin-anchor slice in three, where expanding Deirdre K. Mulligan turned up three affiliation claims and a government title that Berkeley's own faculty page contradicts. Needing the curator's attention are the UK AISI report on Mythos, which now reaches the vault only as a quotation inside a critic's essay and is the clear next capture; Microsoft's MAI Code of Conduct, published today for public consultation and the one deferred item with a known date; the still-undecided placement of the eleven California child-safety bills, now on its fourth deferral; the nine slice-4 single-source pages sharing a 2026-06-06 date, which suggests the three long-deferred slice-3 stragglers are the visible edge of one import rather than three separate gaps; and a CourtListener connector unauthenticated for the eleventh consecutive run.

Categories: Briefings | briefing | gap-scan