Gap Scan — 2026-08-26

Scanned

Recent window: 4 dev-log files (2026-08-24 08:11 and 22:05; 2026-08-25 08:11 and 22:05), 61 pages with last_updated in the last 48 hours. Rotation slice: 13 — sources/ Q–Z (137 pages) + the Raw Sources/ orphan check.

Mechanical scans: bin/lint-scan.py, bin/primary-doc-scan.py --days 3 and --days 4, and bin/ingest-backlog.py. Candidates found: 26. Actioned: 6.

The broken-link lane is empty for the second consecutive run

Wiki/_meta/lint-report.md reports 215 distinct broken targets across 215 refs — 0 referenced 2+ times, unchanged from 2026-08-25. Under the step-3 rule the referenced-once long tail is excluded by design, so the lane generated zero candidates again. Every candidate this run came from the recent-thread scan, the dangling-reference scan, the rotation slice and the deferred backlog. Two of the three pages created here nonetheless cleared broken targets, which is recorded under Link aliases fixed below.

Gaps actioned (6 of 26 found)

New pages created (live)

• Skild AI — score 5 (live thread +3, and a broken [[companies/skild-ai]] target on Google DeepMind +2 on the compute/frontier-capabilities read). Skild was named on two content pages with no page of its own, and released S1 on August 18. Built from the company's own S1 post, TechCrunch's account of the Series C, Sequoia's investment note, Deepak Pathak's CMU faculty page and the University of Maryland's account of the SoftBank round — not from the digest. Carries a valuation Snapshot (>$14B post-money on the January 14, 2026 Series C; ~$4.5B summer 2025; $1.5B on the July 2024 Series A), the founder and business-model record, and both models. 1,180 words, confidence: medium, 5 sources. Depends on queued source: the S1 post, with a provenance note in place.

• U.S. Army (AI Deployer) — score 5 (deferred from 2026-08-25 and re-entering scoring per step 3, +3; government AI procurement core area +2). Gap type 3, and structural rather than transient: the Army is named across nine content pages and holds the largest single AI-related service contract in the wiki's record, while government/ held air-force, cdao and dod and no Army page. Built from AI2C's own designated-laboratory description and Broad Agency Announcement on t2.army.mil and sam.gov, plus the Army-specific facts already sourced on DOD — Department of Defense (AI Deployer), each carried across with its citation. Covers the AI2C mandate and its four key tasks, the Pittsburgh siting on Carnegie Mellon foundations, CamoGPT, the Palantir enterprise agreement and the Raft award, ground autonomy, Project Griffin, the July 2026 token-depletion email, and the four procurement routes. 1,010 words, confidence: medium, 8 sources.

• In-Context Learning — score 4 (live thread +3, gap type 4 +1). The concept had no anchor while the wiki carried GEN-1.5: Embodied Foundation Models are One-Shot Learners as an ingested foundational source turning entirely on it, and named it across GEN-1.5, Generalist and Emergent capabilities — a source with no downstream connectivity. The page defines the capability, dates it to Brown et al. (2020) as the Skild post does, sets the two competing August 2026 embodied accounts side by side in a comparison table with an explicit note that their denominators differ (task success against per-step success), and treats the fine-tuning boundary. 1,290 words, confidence: medium, 5 sources.

Pages expanded (live)

None as a distinct action. The run's budget went to three new pages, their graph wiring and three queued primary documents. The slice-13 finding is a routing decision — see below.

Graph wiring

Nine inbound links added across nine existing pages so the new pages are not orphans and named-but-unlinked references resolve: Multimodality → companies/skild-ai; Defense / Military — AI Deployment, DOD — Department of Defense (AI Deployer) (twice) and U.S. Air Force (AI Deployer) → government/army; Emergent capabilities, GEN-1.5 → concepts/in-context-learning; Carnegie Mellon University → both companies/skild-ai and government/army. Four pages gained substantive additions rather than a link alone: AI Robotics and Generalist gained the S1 counter-account with the non-comparability of the two figure sets stated, GEN-1.5 gained Skild's opposing position on whether in-context learning was trained for, and Emergent capabilities now frames its lead instance as two competing claims rather than one.

Link aliases fixed

None as rewrites, but two broken targets cleared by creation — the same pattern as entities/carnegie-mellon on 2026-08-24 and entities/johns-hopkins on 2026-08-20. [[companies/skild-ai]], referenced once from Google DeepMind, and the newly linked [[government/army]] were not aliases of existing pages and no rewrite could have fixed them. This is the third confirmation in six days that the single-reference tail contains genuine missing-page candidates alongside the author surnames and institution shorthand, which is worth stating because the step-3 rule excludes that tail from candidate generation by design and will keep doing so.

Queued — foundational sources

All three are gap type 9 (uncaptured primary document), each scoring 6 (live thread +3, primary document behind a last-48h digest item +3). All three came from the ## Primary documents sections that schema v4.15 added to the digests. Not ingested.

• OpenAI, "Jalapeño's first results show industry-leading speed and efficiency in AI inference" (2026-08-25). Named with its URL in the 2026-08-25 evening digest; no raw file. Jalapeño is named on four content pages that reach the benchmark result only through trade press. Verified: fetched live from openai.com with cache bypassed, figures corroborated by The Verge and TechCrunch of the same date, two distinctive passages confirmed. Saved: Raw Sources/Jalapeno's First Results Show Industry-Leading Speed and Efficiency in AI Inference (OpenAI, 2026-08-25).md, with the complete numeric appendix. Queued: INGEST-openai-jalapeno-first-results-2026.md.

• Skild AI, "Introducing S1: In-Context Learning for Robotics" (2026-08-18). Named in the same digest, which records that the item is itself the primary document; no raw file. Verified: fetched live from the developer's own domain; the post's own BibTeX block corroborates title, attribution and month, and Pathak's CMU page independently confirms the company and his role. Saved: Raw Sources/Introducing S1 - In-Context Learning for Robotics (Skild AI, 2026-08-18).md. Queued: INGEST-skild-s1-in-context-learning-2026.md.

• OpenAI, "Disrupting a new covert influence campaign from Russia" (2026-08-25). The digest listed this one as unlocated — precisely the case v4.15 added the section to catch — and carried the whole item through CNBC. Found in one domain-restricted search. Verified: fetched live from openai.com; operation, actor origin, platforms and impact assessment corroborated by CNBC. Saved: Raw Sources/Disrupting a New Covert Influence Campaign from Russia (OpenAI, 2026-08-25).md. Queued: INGEST-openai-russia-influence-campaign-2026.md.

Each verification record carries the provenance findings the fold needs. Three are worth naming here because they correct or extend the digest rather than merely supporting it:

1. The Jalapeño announcement date. The morning digest says October 2025, the evening digest reports The Verge dating the first reveal to June 2026, and the two are recorded as unreconciled. The post itself gives no date. The existing sourced record on OpenAI and Broadcom — the OpenAI–Broadcom unveiling of June 24, 2026 — settles it, and the ingest task says to follow that rather than either digest line.
2. The Jalapeño ratios are power-normalized. They compare a 700 W part against 1,200 W and 1,400 W parts, and OpenAI states the normalization explicitly. They are not per-chip ratios, and the comparison systems are named only in the appendix.
3. The Russia report's impact rating is Category Three, not "limited." OpenAI assesses the operation at the lower end of the Brookings Breakout Scale's Category Three — multiple platforms with some indications of breakout to authentic audiences. The digest records only "limited," which understates the report's own calibration.

Authenticity-verification failures

None. All three documents came from their publisher's own live domain, in full body text rather than excerpt, with title, attribution and date corroborated independently.

Two provenance limits are recorded rather than suppressed:

• Firecrawl failed partway through the run. The first six scrape and search calls succeeded; every call after that returned "The connector's server returned an error," including two retries. This is the same failure the 2026-08-24 22:05 digest and the 2026-08-25 gap scan both recorded, now on its third consecutive day. All three queued captures were taken before the failure, so no queued source depends on a fallback path this run.
• One source was sought and not obtained. Anthropic's "In-context Learning and Induction Heads" (Transformer Circuits, 2022) is the standard mechanistic-interpretability reference for the concept. Its existence and canonical URL were confirmed, but the body could not be read — firecrawl was down and the web_fetch fallback returned 124,000 characters on a single line, which neither Read's offset/limit chunking nor the sandbox could slice, since the fetch lands on a host path outside the sandbox mounts. In-Context Learning therefore names the paper and draws no claim from it, and records the absence under ## Open questions. It goes to the deferred backlog as a type-1 candidate rather than being written from memory.

Slice-13 findings

The Raw Sources/ orphan check is clean, and a hand-rolled version of it is not. bin/ingest-backlog.py reports 820 raw markdown files complete, 0 unclassified, 0 foundational without a sources/ page, and 1 non-markdown raw unconverted — the Leo XIV encyclical, structural since v4.10. A hand-written title-matching check run first reported 17 orphans; all 17 were false positives, ingested under sources/ slugs that do not contain the raw file's title string, several defeated by curly apostrophes. This is the same lesson the skill already records for the broken-link scan — "do not write your own scan" — and it now applies to the orphan check as well: bin/ingest-backlog.py is authoritative and should be the only instrument used for it.

Slice-13 thin anchors route to source-robustness-check, as slices 11 and 12 did. Ranked by in-degree against body length: TAKE IT DOWN Act — Source Summary (in-degree 97, 4,486 body chars), Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (53, 4,169), Singapore Model AI Governance Framework for Generative AI (2024) (37, 4,393), The Simple Macroeconomics of AI (29, 3,762), Superintelligence Strategy (Hendrycks, Schmidt, Wang) (26, 3,316), State AG AI Guidances (CA, NJ, MA, OR) (20, 4,362), Sen. Warren Letter to OpenAI (2026-01-28) (20, 3,666), Stanford GSB — Measuring Perceived Slant in LLMs (Westwood, Grimmer, Hall) (18, 3,895). Three consecutive rotation slices have now produced the same routing and nothing has been actioned on any of them. Twenty-one source-robustness-check candidates are outstanding across slices 11, 12 and 13, led by California SB 53 — Transparency in Frontier AI Act at in-degree 254. That skill needs its own scheduled pass; gap-identifier's expand lane is for non-source anchors and cannot absorb this.

Tooling finding: primary-doc-scan.py cannot see the section v4.15 added

All three sources queued this run are gap type 9. bin/primary-doc-scan.py --days 3 flagged none of them, and reported lane 1 clean — "cited primary-document URLs: 3, missing from Raw Sources: 0" — both before and after the run.

The cause is structural, not a threshold. The script re-derives primary documents from item body text using a PRIMARY_HOST regex covering DocumentCloud, CourtListener, regulations.gov, federalregister.gov, congress.gov, govinfo.gov, whitehouse.gov, court domains, named state legislature domains, eur-lex, arxiv.org, and any .pdf. A company-domain post is outside all of them, so openai.com/index/jalapeno-first-results/ and skild.ai/blogs/s1 are invisible to lane 1 even though the digest names both under ## Primary documents with their URLs. The third document had no URL at all — the digest marked it "unlocated" — so no host-based lane could reach it.

The ## Primary documents section that v4.15 added is the authoritative, human-curated list of what a digest rests on, and the scanner does not read it. Parsing that section directly, and treating an unlocated entry as a finding rather than an absence, would have flagged all three. This is a bin/ change and a candidate schema note, so under the ships-alone rule it is reported here rather than made: the scan proceeded by reading the ## Primary documents sections by hand, which is what found the three.

A second, smaller point: the Alabama Attorney General's press release announcing the OpenAI investigation is named with its URL in the same digest and has no raw file, but press-release carries the schema's supporting default. It was not queued. Whether a state law-enforcement office's statement of its own action should keep that default is the same question needs-review/2026-08-23-ap-uber-decision-not-published.md raises for a supervisory authority's decision announcement, and it is now open on two instruments.

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

• Anthropic, "In-context Learning and Induction Heads" (Transformer Circuits, 2022) — new this run, gap type 1, score 5. Named on In-Context Learning with no claim drawn from it. Retrieval failed for the reasons above; retry when firecrawl is responsive.
• Twenty-one source-robustness-check candidates across slices 11–13, above, led by California SB 53 — Transparency in Frontier AI Act (in-degree 254) and TAKE IT DOWN Act — Source Summary (97). Third consecutive slice with the same routing.
• Live-thread items named in the window with no page (score 3 each). Gemini Enterprise for Legal, announced August 25 with Thomson Reuters, Harvey and Legora as connectors — "Gemini Enterprise" is named on five content pages; this is a product fold for Legal Services — AI Deployment and Google DeepMind rather than a page. The Australian Recording Industry Association's revised Charts Code of Practice and the IFPI chart-eligibility principles behind it, both listed as unlocated primary documents; a standards/ page is plausible once the instruments are in hand. TeamT5, the Taiwanese research firm behind the DeepSeek attack-volume finding, named on two pages. Chris Malone's departure as OpenAI head of data centers, which is a fold for OpenAI.
• Two live-thread items examined and declined. Alice, the Israeli AI-security startup reporting a $140 million round at a valuation near $1 billion — zero prior mentions anywhere, and only the first two paragraphs were retrievable behind a subscription wall. The EPA public-comment rule for data-center air permits, which the digest itself files under "Reported but unverified" because nytimes.com refuses automated retrieval; a page built on a headline would fail the quality gate. Both are recorded so they are not re-raised as oversights. government/epa remains a plausible future gap on the broader record — EPA is named on six content pages — but not on this evidence.
• companies/marvell — examined and closed, not deferred a sixth time. Carried since 2026-08-21 and named structural. Reading where the four mentions sit settles it: two are the same June 2026 chip-selloff datapoint on Nvidia & TSMC — AI Compute Infrastructure and index.md, one is Overview's summary of that selloff, and one is the August 18 Google warrant on Google DeepMind. These are illustrative datapoints in other subjects' narratives, not a subject tracked here. Same reasoning applied to companies/uber on 2026-08-23 and companies/tiktok on 2026-08-25. Removed from the backlog.
• legislation/frontier-act — examined and declined; the page exists under another slug. "FRONTIER Act" is named on nine content pages, six of them on Independent Verification Organizations (IVOs) alone, which reads mechanically as a large missing-legislation gap. Alias resolution finds it is Frontier Act / Great American AI Act (Obernolte–Trahan), whose title is already "Frontier Act / Great American AI Act (Obernolte–Trahan)" and which carries the introduced bill in its own section. The August 25 Subramanyam amendment item had already been folded there by the nightly run. The residual issue is that the slug names only the June discussion draft while the page's subject has moved to the introduced bill — a rename, which is supersede's lane, not a gap.
• Carried unchanged. models/deepseek-v3 (in-degree 49 on 608 words, routed to model-robustness-check on 2026-08-22, still unaddressed); the White House release of 2026-07 from the primary-document baseline, instrument still unidentified after seven runs; the remaining slice-9 thin model anchors; the slice-10 stale analysis pages; Amir Salek; the International Catholic Legislators Network; the concept targets parked under needs-review as coined-term and overlap questions.
• The 28 date-stamped headers in lint-report.md (score 1). Unchanged for ten cycles. The finding is unchanged in substance and remains structural: the check does not distinguish legitimate procedural histories from timeline dumps, so the number cannot be driven to zero. It should be fixed in bin/lint-scan.py by excluding litigation/ procedural-history sections rather than re-listed each run.

Notes

• The 2026-08-25 22:05 digest is still unprocessed (unprocessed_devlog=1). Nothing in this run folded a digest item into a page: the three new pages were built from primary sources, and two of the three were standing gaps before that digest existed. The nightly developments-log run should update Skild AI and AI Robotics rather than create anything new for the S1 item.
• Two dedupe drops, both correct. The Stanford HAI data-broker report and the Annenberg W38 survey topline, both named as primary documents in the 2026-08-25 morning digest — raw files and sources/ pages already exist for both, as does one for the Unit 42 report behind the third item.
• No broken outbound links introduced. Every wikilink on the three new pages and the nine edited pages was resolved against the full page set after writing; zero unresolved. One was caught and fixed mid-run: the provenance note on Skild AI used a bracketed sources/ path as illustrative text, which the linter reads as a link — the identical error caught on 2026-08-23, rewritten as prose both times.
• Two frontmatter errors caught and fixed before close. The Skild raw file was written with source_type: technical-report, which is not in the schema enum; the Jalapeño raw file with source_type: press-release, which is in the enum but carries the supporting default and would misclassify a document the wiki cites by number. Both normalized to report, matching the two directly comparable precedents, sources/gen-1-5-one-shot-learners and sources/thomson-reuters-thomson-model-benchmarks. The enum-violation count across Raw Sources/ is back to zero.
• Style scan clean, first pass. All twelve touched pages checked programmatically against the banned-header, hyperbole, self-reference, standing-Predictions and date-stamped-header lists; zero hits attributable to this run. Wiki-wide after: banned_headers 0, self_refs 0, predictions 0, header_hype 0, date_headers 28 — none added.
• Revision fidelity. All nine edits to existing pages were additive — a wikilink wrapped around existing text, a paragraph appended, or a Relationships line added. No page was rewritten, so no _revision-backups/ copy was required and no citation or fact could be dropped. last_updated bumped on the six pages whose content changed; sources_count incremented on the four that gained a citation.
• Confidence set honestly. All three new pages are confidence: medium. None met the three-independent-source bar; Skild AI and In-Context Learning additionally depend on a source not yet ingested and cite it inline pending the fold.
• Cap respected. Six gaps actioned against a 6–10 cap; 26 candidates found. Seven web pulls and two searches before firecrawl failed, one web_fetch fallback attempt, no crawls.
• Dashboard rebuilt. bin/build-dashboard.py wrote 2,029 article pages.

One-line summary

Three pages created live — Skild AI, U.S. Army (AI Deployer) and In-Context Learning — and three primary documents verified and queued, one of which the digest had marked unlocated; the curator's attention is wanted on two mechanical findings, that bin/primary-doc-scan.py does not read the ## Primary documents section v4.15 added and missed all three of this run's queued documents, and that twenty-one source-robustness-check candidates are now outstanding across three consecutive rotation slices with none actioned.