AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-09-12

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

Contents

Scanned

Recent window: 4 New Developments Log/ files (2026-09-10 08:05 and 22:05; 2026-09-11 08:05 and 22:05) and 60+ wiki pages with last_updated inside the last 48 hours. Rotation slice: 2 — concepts/ G–O (44 pages).

bin/lint-scan.py: broken_distinct=212, of which 0 referenced 2+ times and 212 referenced once. banned_headers=0, predictions=0, header_hype=0, date_headers=26, self_refs=6, unprocessed_devlog=1.

bin/primary-doc-scan.py --days 3: 6 digests, 8 cited primary-document URLs, 4 with no raw file, 1 possible match under another address, 3 second-lane keyword hits.

23 candidates survived deduplication. 7 actioned; 16 deferred.

Gaps actioned (7 of 23 found)

New pages created (live)

None. No gap this run met the AUTO-new-page conditions. The three strongest new-page candidates from the recent window — companies/fluidstack, companies/stepfun and entities/tomorrow-together, all deferred by earlier runs — were found to already exist, created by the nightly developments-log fold since the 2026-09-11 scan. They are closed, not deferred. The remaining candidates were either page-update work owned by that same fold, or below the quality gate (see Deferred backlog).

Pages expanded (live)

The rotation slice returned zero thin anchors (no page under 350 words at in-degree ≥6) and only 3 of 44 pages lacking ## Relationships. All three were wired this run — the slice is now complete on that measure.

  • Open-Source AI / Open-Weight Models — score 4 (in-degree 92, the highest in the slice lacking the section; open-source core +2). 1,609 words on 15 sources, and no typed outbound edges at all. Added 11 typed links. Three carry a claim rather than a bare pointer: contradicts: [[concepts/export-controls-ai]], because weights published on the internet pass through no hardware chokepoint and the open-weight case is the standing limit on compute-centric control; contradicts: [[concepts/compute-governance]], the same objection at the level of the governance family, reciprocating the edge that page added on 2026-09-11; and both instance-of: and contradicts: to Meta AI, qualified in place because Meta is simultaneously the largest-scale open-weight case and the clearest case of an open-weight commitment narrowing — reciprocating the pair of edges Meta AI added on 2026-09-10. last_updated 2026-08-30 → 2026-09-12. No prose sentence changed; the existing ## See also and ## Key sources sections were left exactly as found.
  • Mechanistic Interpretability — score 4 (in-degree 52; safety/alignment core +2). 1,871 words on 14 sources, no typed edges. Added 12 typed links tracing the program's position in the safety chain: contradicts: [[concepts/ai-scheming]], qualified in place as a disagreement about whether detection keeps pace rather than about whether scheming occurs; supports: edges to Deceptive Alignment, Alignment Faking and Emergent Misalignment; depends-on: [[concepts/instrumental-convergence]]; and contradicts: [[concepts/monitorability-tax]], qualified as a disagreement about the durability of legible internals rather than their value. A deployed-by: line names Anthropic and Google DeepMind as the two frontier developers running the largest named programs. last_updated 2026-08-27 → 2026-09-12. The existing ## Related concepts section was left intact.
  • General-Purpose AI (GPAI) — score 4 (in-degree 41; EU AI regulation core +2). The stalest page in the slice's top five at last_updated: 2026-07-18, 1,066 words on 8 sources, no typed edges. Added 12 typed links. Two carry a claim: contradicts: [[concepts/open-source-ai]], because an open-weight model below 10^25 FLOP can carry capabilities the systemic-risk tier was meant to capture, and contradicts: [[concepts/distillation]], the same objection from the other direction — capability transfer out of a regulated model passes no compute threshold. Both tensions were already argued in the page's own prose and had never been declared as edges. last_updated bumped. No prose changed.

confidence and sources_count were not changed on any of the three. Graph wiring adds no evidence.

Zero, and the standing zero is now better established than it was. lint-scan.py reports 0 broken targets referenced 2+ times — the candidate pool the skill directs this lane at is empty, not merely unpromising. The 212 single-reference targets are the long tail the skill excludes from re-audit by rule. This is the fifth consecutive scan to find no alias work in that lane.

The one genuine alias defect outstanding — two wikilinks reading [[entities/sen-warren|Bernie Sanders]], invisible to lint-scan.py because they resolve cleanly — is closed. It was repaired by the ingest-reflect cycle on 2026-09-10 and retargeted to Bernie Sanders; the only surviving instance is inside Wiki/_meta/_revision-backups/, which is an archived copy and not a content page. The 2026-09-11 report's instruction to action it this run was already satisfied before this run began.

Queued — foundational sources

  • Healey-Driscoll Administration Statement of Expectations for Responsible Data Center Development and Operations in Massachusetts — score 8 (uncaptured primary document +3, live thread +3, US AI regulation core +2). This closes the item the 2026-09-11 report named "the strongest single capture for the next run," and it was load-bearing: Massachusetts Executive Order 658 — Responsible Data Center Development and Operations documents an order whose Section 1 conditions every covered permit on conformance with this Framework, which the vault did not hold. Verified on mass.gov, the Commonwealth's own domain, HTTP 200, with mg_organization naming both the Executive Office of Energy and Environmental Affairs and the Healey–Driscoll office, and independently corroborated by EO 658 itself, already in Raw Sources/, which incorporates it by reference. Captured complete — Overview, all eight numbered expectation areas, Equity in Large Loads and Energy Use, and Expectations and Guidance in Practice. Saved: Raw Sources/Healey-Driscoll Administration Statement of Expectations for Responsible Data Center Development and Operations in Massachusetts (2026-06-25).md. Queued: INGEST-massachusetts-statement-of-expectations-data-centers-2026.md.
  • Declaration of 25 Fields Medallists, "A Severe Misalignment of AI in Mathematics" — score 8 (live +3, primary document +3, frontier-model core +2). Named in the ## Primary documents section of the 2026-09-11 22:05 digest, whose Frontier Models item rests on it entirely through Terence Tao's blog post; the vault held no copy. Verified on mathandai.org, the signatories' own site, HTTP 200, corroborated by Tao's post under the identical title on his own blog. The signature block was enumerated rather than assumed: exactly 25 names, each with a Fields Medal year. Captured complete. Saved: Raw Sources/Declaration of 25 Fields Medallists - A Severe Misalignment of AI in Mathematics (Math and AI, 2026-09-11).md. Queued: INGEST-fields-medallists-math-ai-declaration-2026.md.
  • In re Stephen D. Aarons, Dispositional Order of Direct Contempt (N.M. Sup. Ct., No. S-1-SC-40845) — score 8 (live +3, primary document +3, core +2). A first-lane primary-doc-scan.py hit from the 2026-09-11 22:05 digest and the instrument its litigation item is about. Captured complete — 4 of 4 pages, all 20 numbered paragraphs. Saved: Raw Sources/In re Stephen D. Aarons - Dispositional Order of Direct Contempt (N.M. Sup. Ct., No. S-1-SC-40845, 2026-09-09).md. Queued: INGEST-nm-aarons-contempt-order-2026.md. Provenance flagged: the copy is a Thomson Reuters legal-documents mirror. The primary host is unavailable for a reason stated on the document's own face — it was not selected for publication in the New Mexico Appellate Reports and is disposed of by nonprecedential order under Rule 12-405(B) NMRA — so this is a structural limit rather than an incomplete search, and it is recorded as such rather than as a clean primary-host capture.

Authenticity-verification failures

  • Liccardo letter to Speaker Mike Johnson (2026-09-11) — score 8, a first-lane primary-doc-scan.py hit. Rejected under protocol step 5: no reachable copy on any trustworthy host. The DocumentCloud URL the digest recorded, and that Axios linked as the copy it obtained, now returns HTTP 404 with a private-content sign-in prompt; the underlying S3 asset returns HTTP 403. A search located the Axios scoop and its CNN and wire pickups, plus one unrelated March 2026 cancel-recess letter on moskowitz.house.gov, but no copy on liccardo.house.gov or the co-signers' sites. The probable reason is that the letter is not final — Axios reports it "was still being circulated as of Friday afternoon," so what circulated was an unsigned draft. Capturing a draft under a title the final letter will also carry would put two different documents in Raw Sources/ under one name. Not saved, not queued. Recorded at Wiki/_meta/queue/gap-scan/needs-review/2026-09-12-liccardo-letter-unreachable.md with the retry condition: a signed release from a member office, at which point the signatory list will be final — the draft's is not, since the digest names four Democrats while at least one outlet described the effort as bipartisan.

Findings that are not gaps

  • The FRONTIER Act resolves by alias and is not a missing page. It is named on nine content pages including Overview and index.md, which reads as a strong missing-page signal. Alias resolution finds Frontier Act / Great American AI Act (Obernolte–Trahan), which carries the introduced bill in a dedicated ## Introduced bill: the Frontier Act (July 2026) section. This was already determined on 2026-08-28. What remains is a curator rename — the slug names a superseded June discussion draft while the page's live subject is the July introduced bill — and the bill text, already recorded as unlocated. Neither is this skill's to decide.
  • legislation/gpai-code-of-practice and sources/gpai-code-of-practice are two pages for one instrument. Found while wiring General-Purpose AI (GPAI). This is the same duplication pattern the 2026-09-08 scan recorded for sources/nist-ai-rmf against standards/nist-ai-rmf, and it is handled the same way: recorded as parallel on the new ## Relationships section rather than merged, since a merge is a curator decision. Two instances of one pattern now; a third would make it structural.
  • Three deferred candidates closed themselves. companies/fluidstack, companies/stepfun and entities/tomorrow-together were all deferred by the 2026-09-10 or 2026-09-11 reports and all three now exist, created by the nightly fold. Recorded so the backlog stays honest, and because it is the second consecutive run in which the fold closed a gap this skill had deferred on the quality gate — which is the division of labour working as intended.

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

New this run

  • Massachusetts Executive Order 654 (score 6), now a confirmed second-order gap. Named twice inside the Statement of Expectations captured today — once in the Overview as one of three things the Framework guides, and once in the Equity in Large Loads section for its directive to the DPU on grid reliability as large commercial customers grow — and once more in EO 658 Section 3. Three references across two held documents and no copy. It is the clear next capture in the Massachusetts thread.
  • Terence Tao, "A Severe Misalignment of AI in Mathematics" (terrytao.wordpress.com, 2026-09-11), score 5. The announcement post behind today's declaration capture, and it carries material the declaration does not: the account of why the signatories forwent a consultative process of the kind used for the Leiden declaration. The 2026-09-11 digest attributes that reasoning to Tao correctly; a future page must not attribute it to the declaration. entities/terence-tao (score 4) is the associated entity gap — he is the declaration's public voice and is named on only two content pages (Astra, AI for Science). Deliberately not created under the quality gate: a page built from one blog post and one signature would be thinner than the declaration ingest will make it.
  • New Mexico Supreme Court order placement (score 4), a decision rather than a gap. The captured order is a sanctions order inside a criminal appeal, not a case the vault would track at Litigation. Whether it lands as a section on Legal Services — AI Deployment plus a sources/ page, and whether a concepts/ page for AI-fabricated filings and professional sanctions is warranted at all, is left to the ingest. The vault currently tracks the pattern in prose with no page behind it.
  • The RubyGems disclosure (score 5) and the two Anthropic threat-report actor findings (score 5 each) — page-update work for the nightly fold on OpenAI and Anthropic, recorded because the 2026-09-11 22:05 digest is unprocessed at scan time. The RubyGems finding is the more consequential: it places an OpenAI agent attack on May 11, 2026, two months before the Hugging Face incident, which moves the start of that thread earlier than any page currently records.

Carried forward

  • The eleven other California child-safety bills signed September 10 (score 5 as a set) — second deferral; the placement question (a page apiece or one page for the September 10 package) is still undecided and is what the cap keeps deferring.
  • Stop Rogue AI Act bill text — the gap is the host, not the task. INGEST-stop-rogue-ai-act-text-2026.md is open and the cited URL is an Inside AI Policy document mirror. Congress.gov should be tried first at ingest time. Recorded against the open task for the second run running rather than raised as a new gap.
  • Third deferral: entities/tech-transparency-project and the TTP report on AI-generated CSAM advertising (score 6, now fourth); the expanded Anthropic Max-tier class action (score 6, fourth, still blocked on an unrecorded court and case number and the CourtListener MCP unauthenticated for the ninth consecutive run); legislation/ pages for the AI OVERWATCH Act, Chip Security Act and MATCH Act (score 5 each, fourth, 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 UK Artificial Superintelligence Bill operative text, the Clement-Jones amendment and the Sanders–Casar press release.
  • entities/adam-thierer (score 3) — fifteenth consecutive deferral, structural. Reported once, not re-listed.
  • Date-stamped headers in lint-report.md — twenty-sixth cycle, structural. Holding at 26. Self-references holding at 6.
  • Carried unchanged. entities/nightingale; legislation/caadca; entities/david-feith and entities/hudson-institute; Marvell; California AB 1856; the Texas OAG / Meta Webster–Mahoney exchange; the two chain-of-thought monitoring papers (arXiv:2507.11473, arXiv:2503.11926); the OpenAI June 2026 Threat Report PDF; the Epoch AI AI-R&D task taxonomy; the ARC Prize Astra evaluation post; Supreme Court of India, Vijay Ghanshyam Gadiya v. Union of India; litigation/worker-info-exchange-v-uber; entities/center-for-shared-ai-prosperity; CMORG "Firm Guidance For Frontier AI"; NCSC frontier-AI guidance; the Minnesota AG Emerging Technology report; ESRB/2026/3; the Governor of the Bank of England's 9 July 2026 letter to Dame Meg Hillier MP.

Slice-2 health, for comparison against prior slices

Slice 2 (concepts/ G–O, 44 pages) is the healthiest slice audited to date on every measure this scan takes.

  • ## Relationships: 3 of 44 lacking, and all three were fixed this run — the slice is now complete. Against 18 of 191 in slice 1, 5 of 168 in slice 0, and 15 of 374 / 9 of 172 / 7 of 149 in the three sources/ slices.
  • Thin anchors: zero. No page in the slice is under 350 words at in-degree ≥6. Slice 1 returned exactly one; every earlier slice returned more.
  • confidence: high on sources_count ≤ 1: 3 of 44, and none at meaningful in-degree — Normative Competence (13), Mathematical Impossibility of Perfect Fairness (9), Mesa-optimization (3). This is the third consecutive slice to show the pattern is a sources/ phenomenon rather than a wiki-wide one, against 102 of 172 and 90 of 149 in slices 12 and 13. The standing recommendation — that a sources/-specific confidence rule belongs in the schema rather than in a nightly report — is now supported by three independent slices and is stated here for the last time as a recommendation rather than a finding.

Verification notes

  • No broken outbound links introduced. All three touched content pages were resolved programmatically against the full page set after writing: zero unresolved on every one. Two candidate targets were caught during drafting and not shipped — concepts/frontier-model-definition and concepts/systemic-risk on General-Purpose AI (GPAI), and concepts/chain-of-thought-monitoring, concepts/sparse-autoencoders and concepts/interpretability on Mechanistic Interpretability — all confirmed non-existent before use and replaced with targets that resolve. The four new queue files and three verification records carry only hedged forward references (entities/raul-torrez, concepts/ai-and-rule-of-law, concepts/ai-and-scientific-integrity), each phrased as a target to decide on; queue files are operational and excluded from the broken-link count.
  • Style scan clean. All three touched pages checked programmatically against the banned-header, hyperbole, self-reference, date-led-header and standing-Predictions patterns: zero hits on each.
  • Revision fidelity. All three edits are purely additive apart from one last_updated value each. No prose sentence, figure, table row, citation or existing section was changed or removed on any page. The ## See also, ## Key sources and ## Related concepts sections on the three pages were left exactly as found, including the bare unqualified wikilinks that are the subject of the standing needs-review note 2026-08-22-ambiguous-bare-wikilinks.md; the new sections carry folder-qualified targets alongside them rather than replacing them, following the precedent set on 2026-09-07.
  • Confidence not inflated anywhere. No confidence or sources_count was changed on any page this run, because graph wiring adds no evidence. No new page was created, so no new-page confidence judgment arose.
  • Source fidelity. Nineteen divergences between primary text and digest were recorded across the three captures, in each raw file's ## Source-fidelity notes. The four that most change a claim: the New Mexico order's $5,000 is payable to the State Bar Client Protection Fund within thirty days, not a fine to the court; that order additionally bars Aarons from appearing before the New Mexico Supreme Court, a sanction no located coverage reports; the Fields Medallists' declaration calls for no regulation, prohibition or pause and states in terms that AI "offers the potential of enhancing and accelerating genuine mathematical study"; and the Massachusetts Framework is guidance in the hortatory voice, whose binding force is entirely derivative from EO 658's incorporation of it.
  • Two dates recorded rather than asserted. Neither the New Mexico order nor the Massachusetts Framework carries a date on its captured face. September 9, 2026 and June 25, 2026 come from coverage and from prior reports respectively; both raw files say so, and both ingest tasks carry the corroboration as a task.
  • Capture completeness stated, not assumed. Massachusetts Framework: Overview, all 8 expectation areas, and both closing sections — complete. Fields declaration: full text plus the full 25-name signature block, enumerated — complete. New Mexico order: 4 of 4 pages, all 20 paragraphs, caption and signature block — complete. Nothing was summarised or trimmed in any of the three.
  • Cap respected. Seven gaps actioned against a 6–10 cap; 23 candidates found after dedupe. Four scrapes and one search; no crawls. Two fetch failures, both on the same rejected document and both recorded above.
  • 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 capture the last two reports called the strongest outstanding one — the Massachusetts Statement of Expectations, on which every EO 658 permit turns — alongside the 25 Fields Medallists' declaration and the New Mexico Supreme Court's AI-fabrication contempt order, and finished the rotation slice's graph wiring on a slice that proved the healthiest yet audited; needing the curator's attention are the Liccardo letter, whose only public copy was withdrawn and which was rejected rather than captured as a draft, Massachusetts Executive Order 654 now confirmed as a second-order gap named three times across two held documents, a second sources/-versus-legislation/ duplicate page for one instrument, and an unauthenticated CourtListener connector for the ninth consecutive run.

Categories: Briefings | briefing | gap-scan