Gap Scan — 2026-09-21

Scanned

Recent window: 4 New Developments Log/ files (2026-09-19 08:09 and 22:05; 2026-09-20 08:05 and 22:08), 506 wiki pages carrying a last_updated in the last 48 hours — almost all of them the evaluations/ build-out shipped under schema v4.16 and v4.17. Rotation slice: 11 — sources/ A–H (427 pages).

Lane results before prioritisation:

• Broken links. bin/lint-scan.py reports 209 distinct broken targets, 0 referenced 2+ times, 209 in the single-reference long tail. The candidate lane is therefore empty by construction, which is the fifth consecutive scan to find it so. Not re-audited, per the settled finding in the skill's autonomy rules.
• Primary documents. bin/primary-doc-scan.py --days 3: 6 digests, 5 cited primary-document URLs of which 2 had no raw file, 0 possible matches under another address, 7 keyword hits in the pre-filter lane.
• Anchor thinness (slice 11). The skill's confidence: medium + sources_count: 1 criterion over-fires on sources/, where one source per page is the normal state — it returned 71 of 427 pages. Re-scored on body length against raw-file length instead, which is the signal that actually distinguishes an underbuilt source page. Top results by that measure: claude-sonnet-46-system-card (487 words against a 29,082-word raw, 1.7%), california-ccpa-regulations (433 against 41,547, 1.0%), gpt-54-thinking-system-card (472 against 2,870, 16%), california-sb-243 (417 against 1,334, in-degree 120). altman-abundant-intelligence was dropped: at 405 words against a 367-word raw the page is longer than its source and is correctly built, not thin.
• Recent thread. Named actors, laws and models from the four digests were checked against the graph. The nightly developments-log runs had already created pages for Hacktron, Nscale, the American AI Security Act, TypeSafe AI, Irregular, Colorado HB 26-1263, CXMT, and both September 18 executive orders. Genuine misses: RADAR/Damo Radar, the Hobbs veto letter, Tech Oversight Project, Corridor, Mike Lawler.

Gaps actioned (7 of 19 found)

New pages created (live)

• OpenRCA — score 6 (in-degree from four content pages, safety/frontier-lab governance core +2, schema v4.17 obligation +1, live thread +3 via the slice-11 card). Gap type 3. OpenRCA is reported in two Anthropic system cards, has a peer-reviewed ICLR 2025 paper and a Microsoft-hosted MIT-licensed repository, and under v4.17 ("a public paper or leaderboard") earns its own page — but it had none, and its scores sat in prose on Claude Opus 4.6 and Claude Sonnet 4.6 where no two could be compared. Built from the OpenReview record, the repository, and both system cards. Carries one fixed-format ## Results track (Overall, eight rows); the per-system decomposition sits outside ## Results as its own section, because its columns are components of the Overall figure rather than a separate track and bin/benchmark_decoder.py correctly rejected them as a track on the first build.

The page records a criticism the vault did not hold: a July 2026 repository issue measuring a zero-information constant-prediction baseline at 7.46% strict accuracy against the paper's headline 11.34% for RCA-agent, with the sign of the difference reversing on two of the four datasets and both scoring zero on the one task shape asking for all three fields. The submitter reproduced the published 11.34% exactly before reporting the baseline, and frames the finding as being about reporting rather than validity. The issue is open and unanswered (Source: github.com).

• RADAR (abdominal CT diagnosis model) — score 5 (live thread +3, uncaptured primary document +3, capped at the gap's own merit). Gap types 3 and 9. Named in two digests and on one content page, with no model page. Two naming and four substance corrections against the coverage the vault had folded:

The authors call the model RADAR; "Damo Radar" is the repository name that coverage adopted. The Science paper published 17 September, not the 18th; the repository's earliest directories date from early July 2026 and carried 45 commits by 21 September, so the September dates mark publication and announcement rather than first code release. On substance, the abstract reports performance "for 18 anatomical structures and 146 imaging findings" (coverage: "close to 150 conditions"), a reader study in which assistance "increased the diagnostic sensitivity of 26 radiologists by ~10%" (coverage: radiologists "reduced missed diagnoses by 10%" and the model "outperformed 23 participants"), no AUC figure at all (coverage: 0.913), and a conclusion that generalist AI "can match human experts" rather than outperform them. The page also records that the repository states its licence inconsistently — Apache 2.0 in its License section and LICENSE file, CC BY-NC-SA 4.0 in its badge — which bears directly on the "open-sourced" framing.

Pages expanded (live)

• Claude Sonnet 4.6 System Card — score 5 (in-degree 56 +2, safety core +2, v4.17 +1). 487 → 2,203 words. The page covered a 29,082-word card in five short sections and named five benchmarks. It now carries the capabilities chapter including the new multilingual section (GMMLU and MILU gaps by model, and the thinking-token efficiency measurement: 1,078 tokens per question for Gemini 3 Pro against 127 for GPT-5.2 Pro), the safeguards chapter, the full alignment assessment with its fourteen summary findings, the internal pilot-monitoring results, agentic safety with the Gray Swan Shade prompt-injection figures, and the RSP determinations.

Two corrections. The page said "For autonomy risks, the card concludes ASL-3 is appropriate and no ASL-4 trigger was reached." The card in fact states that Anthropic proactively implemented AI R&D-4 safety measures despite not judging the threshold crossed, and that "it is plausible that models equipped with highly effective scaffolding may be close to this AI R&D-4 threshold." The page also said that for cyber risks "no critical threshold was crossed"; the RSP defines no cyber threshold at any AI Safety Level, and the card says the model "is close to saturating our current cyber evaluations." Also newly recorded: Anthropic states it "was not able to arrange for an in-depth alignment-focused third-party assessment" for this model. Per v4.17 the card's individual scores stay on their evaluations/ pages, which the 48-hour build-out had already populated; the source page links the set rather than duplicating rows. Original backed up to _meta/_revision-backups/sources/claude-sonnet-46-system-card.2026-09-21.md.

• California SB 243 — Companion Chatbots — score 4 (in-degree 120 +2, US AI regulation core +2). 417 → 1,216 words. The highest in-degree page in the slice summarised the statute in a five-bullet list. It now carries the Legislative Counsel's Digest framing, the full § 22601 definitions including all three exclusions and the definition of "video game" they turn on, and the operative provisions by section: the three-hour break notification for known minors, the sexually-explicit-material provision, the requirement that a chatbot not engage at all absent a published crisis protocol, the July 1, 2027 reporting start with its three enumerated items and the bar on user identifiers, the evidence-based-measurement requirement, and the private right of action at the greater of actual damages or $1,000 per violation plus attorney's fees — none of which the page previously stated. Codification at Bus. & Prof. Code §§ 22601–22606 added. Every prior fact and citation preserved; original backed up.

• Healthcare — AI Deployment — graph wiring for the new model page, plus the primary-source correction above set alongside the coverage-derived figures it already carried. No prior fact removed.

Link aliases fixed

None. Zero broken targets are referenced twice or more, so there is no candidate to alias-resolve. This is the fifth consecutive scan to report zero; the remaining 209 are the single-reference long tail of author surnames, institution shorthand and deliberate forward references, and were not re-audited.

Queued — foundational sources

• Tony Fabrizio (Fabrizio Lee), "Neutralizing Political Attacks on AI Data-Center Development Through Enforceable Conditions" (2026-09-14) — score 8, the run's highest. Gap type 9, precise lane. The 2026-09-20 morning digest quoted the memo's opening sentence and flagged the document in its own ## Primary documents section as an INGEST candidate. Verified: DocumentCloud (secondary host; a confidential consulting memorandum has no primary publisher, same handling as the NRSC "Ohio Data Center Risk" memo under schema v4.15), corroborated against WIRED's verbatim quotation and the letterhead date. Full text, not an excerpt. Saved: Raw Sources/Fabrizio Lee - Neutralizing Political Attacks on AI Data-Center Development Through Enforceable Conditions (2026-09-14).md. Queued: INGEST-fabrizio-data-center-messaging-memo-2026.md.

Two discrepancies recorded for the ingest. The digest describes the memo as "following an August memorandum from the NRSC"; the memo contains no reference to that document, and the sequencing is the coverage's. And the memo's closing section states three limits on its own survey — it does not test vote share, the three conditions were tested as a package so their independent effects are unknown, and the contrast "does not guarantee the issue disappears" — none of which reached the digest. The underlying survey's sponsor, field dates and sample size are not named anywhere in the document.

• Zhang et al., "An expert-level generalist AI for abdominal CT diagnosis," Science 393(6817):eaec6129 (2026-09-17) — score 5. Gap type 1, and the dependency of RADAR (abdominal CT diagnosis model), which carries a provenance note pointing at it. Verified: science.org DOI landing page, HTTP 200, DOI resolves, 40-author list and volume/issue/article number match the BibTeX the authors publish in their own repository. No raw file saved — the full text is paywalled; abstract, editorial summary and author list captured, PDF to be fetched fresh at ingest. Queued: INGEST-radar-science-abdominal-ct-2026.md, which carries the four coverage-versus-abstract divergences above.

• Katie Hobbs, veto letter for Arizona HB 2311 (2026-06-19) — score 8. Gap type 9. A primary-document URL cited in the 2026-09-20 22:08 digest with no raw file; the string "Hobbs" appeared on no content page. NPR's investigation reports that near-identical chatbot-bill exemption language appeared in at least ten states and that Arizona's was vetoed; this letter is the only primary statement of why one was rejected, objecting to a damages cap "that amounts to a drop in the bucket for large corporations," a bar on families "bringing their own lawsuits," and constraints on state enforcement she calls "unprecedented." Verified: azleg.gov, the Arizona State Legislature's own domain, HTTP 200, URL path encoding the 57th Legislature 2nd Regular Session. Complete one-page letter. Saved: Raw Sources/Arizona Governor Katie Hobbs - Veto Letter, HB 2311 Artificial Intelligence Service Disclosures (2026-06-19).md. Queued: INGEST-hobbs-veto-letter-hb2311-2026.md. The letter names neither Google nor industry drafting; any link to the NPR allegations is the reporting's, not the letter's.

Authenticity-verification failures

None. Four documents were pulled; four verified. One (the Science paper) was verified at its canonical host but not saved, because the full text is paywalled — recorded as a deliberate queue-without-raw rather than a failure, per the skill's allowance for a large source best fetched fresh at ingest time.

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

• California CCPA Regulations (Title 11, Division 6) — 433 words against a 41,547-word raw (1.0%, the worst ratio in the slice), in-degree 28. Score 4. The largest remaining slice-11 expansion; deferred on size, not on merit. Second deferral.
• GPT-5.4 Thinking System Card — 472 words against a 2,870-word raw, in-degree 40. Score 4. Same class as the Sonnet card handled this run.
• America's AI Action Plan (in-degree 300, 942 words) and California SB 53 — Transparency in Frontier AI Act (in-degree 277, 678 words) — score 4 each. The two highest-in-degree pages in the entire wiki. Their raw files were not measured this run; whether they are genuinely underbuilt or merely short relative to their sources needs the ratio check before either is actioned.
• entities/tech-oversight-project — score 2. Named on four content pages (The OpenAI Files, two ARI coalition-letter pages, Americans for Responsible Innovation (ARI)) and quoted in the 2026-09-20 NPR item, with no page. Genuine gap type 3, below the cap on score alone.
• Arizona HB 2311 bill text — score 4. The vetoed bill itself has no raw file, so the veto letter's characterisation of its damages cap and enforcement provisions cannot be checked against the text. Noted in the ingest task.
• companies/corridor and entities/mike-lawler — score 3 and 2. Jack Cable's company is quoted in the Gemini containment item and named on two content pages; Lawler co-sponsors the American AI Security Act with Josh Gottheimer, who has a page.
• Named-lane pre-filter items not actioned — the Brent Hecht deposition material in the NYT case, the DOJ 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; the September 19 Axios item appears to describe the same filing and this should be confirmed rather than re-pulled), the IMF Dublin background note, France's UN Security Council concept note (reported as not public), the Bowling Green State University Ohio poll, and the NYT/Siena crosstabs.

Structural note. Wiki/_meta/queue/gap-scan/needs-review/ holds 14 dashboard-rebuild-failed notes from June 2026 and 22 curator decisions, the oldest from 2026-07-09. The decision notes are questions waiting on the curator and are correctly excluded from re-raising, but the June rebuild-failure notes are stale artifacts of a fixed problem and are now noise in a directory the dedupe step reads every run. Worth a sweep.

One-line summary

Seven gaps actioned: two pages created (OpenRCA, RADAR (abdominal CT diagnosis model)), three expanded, three foundational sources queued for the curator; the run's substantive finding is that three separate pages carried a coverage claim its primary source does not support — the Sonnet 4.6 card's RSP determinations, RADAR's reader-study and accuracy figures, and the Fabrizio memo's own stated limits.