Gap Scan — 2026-08-10

Scanned

Recent window: 4 New Developments Log/ files (2026-08-08 08:10 and 22:11; 2026-08-09 08:09 and 22:10) and 51 wiki pages edited in the last 48 hours. Rotation slice: 11 — sources/ A–H (325 pages).

Broken-link baseline at the start of the run: 270 distinct broken targets across 304 instances, mainspace only. Two scanner corrections were made before scoring, both of which had been inflating the candidate list on previous runs:

• Escaped-pipe artifact. Targets captured with a trailing backslash (companies/anthropic\, models/gpt-55\) come from [[page\|Alias]] inside markdown tables and are not broken. Stripping the trailing backslash removed 66 phantom targets. This is the same defect flagged as the bin/lint-scan.py "escaped-pipe target-capture bug" on each of the last ten runs; the fix is one .rstrip("\\") in the target-capture regex and is worth applying to bin/lint-scan.py itself.
• Basename in-degree conflates folders. Counting in-degree by basename merges sources/california-sb-243 with legislation/california-sb-243. Six of the ten apparent thin anchors in this slice dissolved once in-degree was computed per qualified target — sources/california-sb-243 has 62 inbound by basename but far fewer as a sources/ target, and its legislation/ counterpart is well built. No sources/ A–H page trips the strict thin-anchor rule (in-degree ≥6 and under 350 words). The nearest misses are recorded in the deferred backlog rather than actioned.

Nine candidates survived alias resolution and deduplication against the open queue, the seven gap-scan reports of 2026-08-03 through 2026-08-09, and existing Raw Sources/ files. Seven were actioned.

Gaps actioned (7 of 9 found)

New pages created (live)

• Behind-the-Meter Power for Data Centers — topical hole. On-site and co-located generation supplying data centers directly is a recurring subject across the last four developments files (Amazon's Pecos County plant, the Lancium and Crusoe campuses, xAI's Southaven turbines, SpaceX's Southaven build-out) and had exactly one passing mention wiki-wide, a parenthetical on Data Center Siting / AI Power Politics about the Talen–Susquehanna structure. Built from five independent sources: Enverus on interconnection-queue timelines (over 2,100 days average, up ~60% since 2017; ~10% of queued capacity built), A&O Shearman on the model's reliability limits and the EU Grids Package, Energy Innovation's Utility Dive opinion on ratepayer and emissions effects (100 GW of planned on-site gas capacity per BNEF, 18% of existing US gas capacity), datacenterHawk on adoption (Foley 2026 survey: 56% of developers exploring on-site generation), and Orrick on network-load classification. confidence: medium. One inbound link added from Data Center Siting / AI Power Politics.
• Raja Krishnamoorthi — broken-link cluster plus core-area gap. Two inbound references from John Moolenaar, one of them a typed co-engaged-with: relationship, both broken. The chair of the House Select Committee on the CCP had a page; the ranking member did not, leaving the Democratic side of the AI chip export-control debate unrepresented. Built from four official House Select Committee (Democrats) press releases and one Legistorm filing notice, including his December 9, 2025 statement on H200 exports and the Krishnamoorthi–Bera No Advanced Chips for the CCP Act. confidence: medium. Both previously broken links now resolve.

Pages expanded (live)

None. The thin-anchor rule found no qualifying page in slice 11 once in-degree was computed correctly — see the scanner correction above. This is the finding, not a shortfall.

Link aliases fixed

• [[sources/when-the-interface-is-neural]] → [[sources/farahany-adv-class-10-3-when-interface-is-neural]] (6 instances across concepts/farahany-cognitive-evidence-spectrum, concepts/cognitive-liberty, sources/farahany-adv-class-10-2-two-courts-one-test-one-thumb, sources/farahany-adv-class-11-1-wall-that-cuts-both-ways, entities/nita-farahany, and the Farahany class sequence in index.md)
• [[entities/spacex]] → [[companies/xai|SpaceX]] (1 instance, sources/ft-ai-revolution-turbocharged-ma), following the convention recorded on U.S. Air Force (AI Deployer) that the merged SpaceXAI entity is tracked under xAI. A separate companies/spacex page was considered and rejected: it would split one organization across two folders, against the no-duplication rule. The [[companies/spacex]] targets that appear in the broken-link scan are all inside _meta/ gap-scan reports, not mainspace.
• [[wikilinks]] → wikilinks (4 instances across concepts/frontier-ai-governance, entities/meta-oversight-board, and the two stub sources/ pages below). These were the literal word inside brackets in provenance prose, not a link target; they were also wiki self-reference under house style §2d, and were removed as such.
• Removed a redundant related: line pointing at a superseded stub from Claude Sonnet 4.6. No fact and no citation lost — the same page already carries the underlying citation in prose at its User wellbeing section.

Two stub sources/ pages resolved and marked superseded

Both were created by the 2026-06-06 gap scan as landing pads for sources believed unidentified. Raw Sources/ held the answer in both cases.

• sources/when-the-interface-is-neural was a duplicate slug for the already-ingested Inside My Advanced Topics Class 10.3: When the Interface Is Neural (Farahany, April 2026) (Nita Farahany, April 2026), which is separately listed in index.md. Set status: superseded, superseded_by the real page.
• sources/protecting-wellbeing-of-users was reserved for Anthropic's "Protecting the wellbeing of our users," saved at Raw Sources/Protecting the wellbeing of our users.md since 2026-04-13 and classified source_class: supporting — which under the classification rules takes no Wiki/sources/ page at all. Set status: superseded.

Removal of both files needs the curator's sign-off and is recorded at Wiki/_meta/queue/gap-scan/needs-review/2026-08-10-two-stub-sources-pages-superseded.md, together with a recommendation to run Migrate-Supporting on the 2,169-word Anthropic raw file, whose account of the self-harm classifier, the Claude.ai support banner, the sycophancy work and the 18+ requirement is unfolded anywhere except two evaluation figures.

Queued — foundational sources

• APA, "Artificial Intelligence and Adolescent Well-being: An APA Health Advisory" (June 3, 2025) — dangling foundational reference, cited twice from AI's Delusional Spirals (and What to Do About Them) — Stanford HAI / Moore et al. (2026), once in prose as the clinical anchor for adolescent AI mental-health risk and once as a typed supports: relationship, with no page behind either. Verified: apa.org (canonical host; 11-page PDF, HTTP 200, contentType: application/pdf), corroborated by the APA landing page, the June 2025 APA press release, and the companion chatbots/wellness-apps advisory which cross-links it by exact title and URL. Saved: Raw Sources/APA Health Advisory on AI and Adolescent Well-being (2025) - primary text.md. Queued: INGEST-apa-health-advisory-ai-adolescent-well-being.md. Verification record: Wiki/_meta/queue/gap-scan/proposed-sources/apa-health-advisory-ai-adolescent-well-being-2025.md.

The existing raw file for this source fails the authenticity protocol. Raw Sources/APA Health Advisory AI Adolescent Well-being 2025.md (created 2026-04-14) would normally have removed this candidate under the "a Raw Sources/ file already exists" dedup rule. Its body states on its face: "Compiled from apa.org public materials. Not a verbatim copy." Protocol step 5 rejects content that is a summary or compilation rather than full text. It also merges two distinct APA advisories — the June 2025 adolescent advisory and the November 2025 chatbots/wellness-apps advisory — under a single published: field carrying both dates. The primary text was pulled alongside it; the ingest task flags the pair for Dedupe-Raw or Merge-Raw resolution. The dedup rule as written treats the presence of a raw file as sufficient, which is what let a non-verbatim compilation sit unchallenged for four months.

Authenticity-verification failures

One, and it was an artifact already inside the vault rather than a fresh pull: the 2026-04-14 APA compilation described above. It is left in place as a research note, not deleted, with the rejection reason recorded in the verification record and the ingest task.

The APA advisory itself passed every step. One weakness is recorded rather than waved through: an APA health advisory carries no DOI, ISBN, or docket number, so the identifier check rests on the canonical apa.org publication path plus the two independent cross-references. That is a weaker identifier link than a DOI-bearing publication would give — the same caveat recorded for the OpenAI newsletter pull on 2026-08-08.

A second, smaller fidelity note: the PDF parser transposed leading characters on several reference-list entries ("4euerriegel, S." for "Feuerriegel, S."). The saved raw file carries the corrected list with an explicit note that individual citations should be checked against the published PDF before quoting, rather than silently presenting repaired text as extracted text.

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

• Near-threshold thin anchors in slice 11 (score 2). sources/emotion-concepts-llm (55 inbound, 436 words) and sources/gpt-54-thinking-system-card (38 inbound, 425 words) sit just above the 350-word floor with high reliance. Both are sources/ pages with a Raw Sources/ file behind them, which makes them source-robustness-check work rather than gap-identifier expansion from supporting web sources.
• Ambiguous bare wikilinks — a systemic finding, not a page gap. Roughly 1,700 link instances use a bare basename that exists in two folders at once: [[eu-ai-act]] (250 instances, resolving to either sources/eu-ai-act or legislation/eu-ai-act), [[americas-ai-action-plan]] (211), [[california-sb-53]] (160), [[colorado-ai-act]] (141), [[eo-14365]] (132), [[techno-federalism]] (95, sources/ vs concepts/), [[nist-ai-rmf]] (94, sources/ vs standards/), [[ai-as-normal-technology]] (93, sources/ vs concepts/), and about fifteen more pairs. These render in Obsidian but each one silently picks a folder, so in-degree, the retrieval index, and the dashboard graph are all computing against a guess. Bulk rewriting is out of scope for a gap scan — each instance needs a judgment about whether the source or the law is meant — but it is the largest single distortion in the link graph and belongs in a scoped lint pass.
• entities/tim-fist and entities/saif-khan (score 1). Both named in the 2026-08-09 evening file as authors of the Institute for Progress argument for pacing automated AI R&D; Fist has one prior mention, Khan none. Below the reference-count threshold, and Institute for Progress (IFP) already exists to carry the position.
• Person pages implied by this window's pulls, none created: Adele Lopez has a page; Jeff Ding, Lucas Hansen (CivAI), Zak Stein (AI Psychological Research Coalition), and Jeffrey Rissman and Eric Gimon (Energy Innovation, now cited on the new behind-the-meter page) do not. All at one mention.
• Organizations named in the window with no page: Trajectory Labs (ran Anthropic's commissioned 720-attempt prompt-injection evaluation), CivAI, the AI Psychological Research Coalition, Cleanview, Pacifico Energy, Energy Innovation. Trajectory Labs is the strongest of these and is likely to recur.
• [[_meta/briefings/weekly-2026-W19]] — resolved as a scanner artifact, closing a nine-run carry. The target file exists at Wiki/_meta/briefings/weekly-2026-W19.md; the scan flagged it only because _meta/ is excluded from the content-slug set. It is not a broken link. What it is instead is two mainspace pages (concepts/ai-macro-prudential-policy and index) linking into the operational _meta/ subtree, which the v4.3 content/operational split discourages. Reframed for the next run rather than carried again as "broken."
• entities/consumer-technology-association — carried since 2026-08-07, still one inbound reference, still below threshold.
• The 2026-06-06 stub-creation cohort. Both stubs resolved today came from that run and shared one failure mode: creating a placeholder from a broken link's slug without first searching Raw Sources/ or checking for an existing ingested page under a different name. Worth a scoped check of whatever else that run created.

One-line summary

Two new pages created live (Behind-the-Meter Power for Data Centers, Raja Krishnamoorthi), eleven broken links repaired, two long-standing stub sources/ pages traced to files that were already in the vault and marked superseded pending the curator's deletion decision, and the APA adolescent-AI health advisory pulled, verified, and queued for ingest after the compilation standing in for it was found to be a non-verbatim summary; broken-link count fell 304 → 297 instances, and two scanner defects that have been inflating the candidate list for ten runs are now documented with fixes.