Gap Scan — 2026-08-08

Scanned

Recent window: 4 New Developments Log/ files (2026-08-06 08:10 and 22:12; 2026-08-07 08:05 and 22:05), 34 items. Rotation slice: 9 — models/ + industries/ (76 pages: 63 models, 13 industries).

Broken-link analysis ran over all 1,774 content pages and 2,526 distinct link targets. The mechanical broken-link count is materially wrong and has been for ten runs. bin/lint-scan.py reports 341 distinct broken targets; corrected for the escaped-pipe capture bug — [[page\|Alias]] yields a target ending in \ — the true figure is 255, and of those only 7 have inbound references from two or more distinct pages. The head of the distribution (companies/anthropic\, companies/openai\, companies/meta\, gpt-53-codex\) is entirely artifact. This run worked from the corrected set.

Gaps actioned (9 of 24 found)

New pages created (live)

• NetChoice — the highest-in-degree genuine missing page in the wiki: named across 27 mainspace pages, with NetChoice v. Bonta (CAADCA litigation) and Moody v. NetChoice both load-bearing on Algorithmic Speech Doctrine, AI and the First Amendment and Content vs Architecture Theory of Social Media Harm, and no organization page behind any of it. Carried on the deferred list from 08-07 with an explicit "create if missing" marker. Built from the association's own about page, litigation index and Connecticut testimony, plus the Justia case summary. confidence: medium, 4 sources.
• Koray Kavukcuoglu — now runs a frontier lab and had no page; named on 4 mainspace pages including Google DeepMind and Demis Hassabis, both of which recorded the August 5 change without an entity behind it. Third run to carry him. Built from Pichai's and Hassabis's own messages plus Fortune and Semafor. confidence: medium, 4 sources.
• Taalas — live thread: AMD announced a definitive agreement to acquire it on August 6, 2026. A hardware maker with a distinct architectural bet (weights etched into fixed silicon) that the compute thread will need to cite by name. Built from AMD's own investor-relations release, Quartz and unite.ai. confidence: medium, 3 sources.
• Lancium — live thread: Nvidia agreed to invest up to $3 billion. Lancium owns the Abilene campus that is the first operational Stargate site, so it sits directly under Stargate Project, and its 4 GW secured / 15 GW queued position is the concrete case for the Texas interconnection freeze. confidence: medium, 2 sources.
• Adam Cassady — confirmed 51–47 on August 7 as US ambassador at large for cyberspace and digital policy, the second person ever to hold the post, and on the record declining to take a position on advanced chip exports to China. confidence: low on a single source, which is the honest rating.

Pages expanded (live)

• Astra — the page ended at "no safety documentation has been published," which stopped being true on August 7. Added a ## Cyber-capability finding and development pause section: OpenAI's statement that it "cannot rule out critical cyber capabilities," the Preparedness Framework trigger, the paused internal activities and isolated testing environments, the White House notification, the contrast with Anthropic's rolled-back RSP pause commitment, and Michael Dalton's Black Hat remarks. Repaired safety_case frontmatter, the lead, and the safety Open-question; added one more. 1 new inline cite, sources_count 10 → 11.
• Legal Services — AI Deployment — slice-9 thin anchor: in-degree 17 against 782 words, untouched since 2026-07-26, and carrying no vendor economics at all. Added ## Vendor scale and funding covering Harvey's proposed $15.5B round, the full valuation sequence from $3B, the ARR series, the investor list, the 44× multiple, deployed scale, and the Clio / Relativity / Legora competitive set. 2 new inline cites, sources_count 7 → 9, 782 → 1,094 words.
• Harvey — expanded to resolve a contradiction the industries/legal edit would otherwise have created. The page's lead read "a September 2025 Fortune profile reported a $100 million valuation," which cannot be reconciled with a $3 billion valuation in February 2025; the $100M figure is ARR. Added ## Snapshot Valuation and Revenue tables (five and three dated rows), corrected the lead, corrected the AmLaw figure, and recorded the two conflicting self-reported scale figures side by side rather than picking one. last_updated 2026-06-06 → 2026-08-08, sources_count 5 → 8. Also removed a wiki self-reference ("The page also lists…") found in passing.

Link aliases fixed

None required. All seven genuinely broken multi-inbound targets resolved to one of three non-gap categories (see Deferred). The 14 [[companies/nvidia]] references flagged mid-run were introduced by this scan's own drafts and were corrected to [[companies/nvidia-tsmc]] before commit; there is no pre-existing cluster.

Queued — foundational sources

• OpenAI Global Affairs, "Making AI Audits and Assessments Work" (2026-08-07) — a named six-principle scheme a frontier developer is proposing for third-party oversight, carried in detail by the 08-07 22:05 digest with no sources/ page and no raw file behind it. Verified: openaiglobalaffairs.substack.com (OpenAI's own global-affairs newsletter; forced live fetch, HTTP 200, article:modified_time 2026-08-07T22:45:34Z; four distinctive passages matched; independently described the same day by the digest; all policy outbound links resolve to openai.com/index/). Saved: Raw Sources/OpenAI Global Affairs - Making AI Audits and Assessments Work (The Prompt, 2026-08-07).md (full post, all three sections). Queued: INGEST-openai-ai-audits-assessments-2026.md. Verification record: queue/gap-scan/proposed-sources/openai-ai-audits-assessments-2026-08-07.md.

Non-gaps confirmed by alias resolution

• concepts/spiralism was scored into the top six and then withdrawn. The August 6 Verge feature made "spiralism" a live thread with 7 mainspace mentions and no page under that name. Alias resolution resolved it to Parasitic AI / Spiral Personas, which defines Spiralism as a term inside Adele Lopez's framework, and the Verge coverage had already been folded into both that page and Adele Lopez on 08-07. Creating a spiralism page would have split a coherent concept across two slugs. Not created; not a gap.
• companies/discovery-loop — deferred on 08-06 as the strongest new-page candidate; created since. Dropped from the list.
• entities/consumer-technology-association — carried from 08-07; still 1 inbound reference, still below threshold.

Authenticity-verification failures

None. One source pulled and queued; it passed every step of the protocol. One weakness is recorded rather than waved through: a company newsletter post carries no external identifier the way a paper or a court filing does, so the identifier check rests on the canonical URL, the publication identity and the openai.com outbound link set. That is noted in the verification record.

Post-run verification pass

An independent fidelity audit was run over all eight touched pages against their live cited sources — roughly 145 discrete claims, sixteen sources re-scraped. No direction reversals, no polarity reversals, no fabricated quotes; every direct quotation checked appears verbatim in its cited source. Date discipline held on four of five event dates. Ten defects were found and all ten are corrected here:

1. A legal holding mischaracterized. NetChoice described Moody v. NetChoice as decided by "the plurality." Kagan delivered the opinion of the Court, joined in full by Roberts, Sotomayor, Kavanaugh and Barrett and in part by Jackson; Thomas, Alito and Gorsuch concurred in the judgment only. Corrected and the line-up recorded.
2. A quotation's scope widened. The same page attributed "a trade association whose members include Facebook and YouTube" to NetChoice alone. The Court's phrase is plural and describes NetChoice and CCIA together. Corrected.
3. Two anonymized actors, both named in the source. The amicus brief was "a suit" dismissed by "a district court" — it is Bogard v. Alphabet, filed July 2, 2026, supporting Google and TikTok. The Connecticut testimony was "Connecticut AI legislation" — it is SB 5, delivered by Patrick Hedger to the Joint Committee on General Law on March 3, 2026. Both corrected.
4. A dropped limb of an argument. The Connecticut testimony raises three defects; the page carried two. The Due Process vagueness objection — to "reasonably foreseeable," "capable of" and "catastrophic risk" — is now recorded.
5. An announcement date written as an effective date. Koray Kavukcuoglu said he "became" SVP "on August 5, 2026." Both Pichai and Hassabis write "will step up"; no effective date was given. Corrected to "was named," with the absence of an effective date stated.
6. An unsourced role change. The same page said he "had previously served" as chief technology officer. Pichai calls him "the current Chief Technology Officer of GDM" and neither source says he vacates it. Corrected.
7. Two more anonymized actors. Oriol Vinyals and Quoc Le left with Dean and Ghemawat and were unnamed. Added, along with Discovery Loop's stated purpose and the Gemini 3.5 Pro / Shazeer / Jumper context Fortune supplies.
8. A citation on the wrong source, and a one-sided compression. Taalas cited AMD's press release for the metal-layer mechanism, which appears in unite.ai, not AMD. Re-cited. The page also carried the model-lock constraint without Taalas's rebuttal (fewer than a handful of 100+ layers change between designs; roughly two months design-to-silicon) and dropped two qualifiers on the throughput figure ("relative to GPU benchmarks"; second-generation silicon moves to 4-bit). All restored.
9. An attribution shifted. Lancium had "the companies saying they would invest up to $500 billion." Reuters attributes the figure to Trump's statement. Corrected.
10. Unsourced precision, an anonymized primary source, and attribution drift on Astra. "Two days earlier at the Black Hat conference" — Axios says only "earlier this week," and Black Hat USA ran August 5–6, so the interval is not resolvable; corrected to "earlier the same week." The security-controls detail comes from an OpenAI blog post published the same day, which the page never mentioned even though an Open-question turns on what OpenAI has put on the record; now named. Dianne Penn's "deliberately more conservative" characterises the company's posture, not the release; corrected, and Axios's description of Mythos as its most cyber-capable model restored. "The Hugging Face exploits attributed to its models in July 2026" carried a month not in the cited source; the qualifier is removed.

Also corrected in the same pass: Harvey's "more than 60 AmLaw 100 firms" was attributed to the company website, which now says 75+ — the live stat block (200,000+ professionals, 2,400+ firms and teams, 70+ countries, 25+ hours saved per month) is now recorded with its own citation alongside the August reporting figure it conflicts with; a self-reported figure recast as third-party reporting was re-attributed to the company; and an inference the wiki was drawing in its own voice about how to read the Fortune headline was replaced with the two dated facts, leaving the reading to the reader. sources_count was over-declared on three new pages (adam-cassady 2→1 with confidence medium→low, taalas 4→3, lancium 3→2) and under-declared on harvey (7→8).

One item found by the audit is not fixed and is carried below: an internal inconsistency between Astra's Open-question 4 and Safety and Alignment in an Era of Long-Horizon Models (OpenAI, July 2026).

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

• Slice-9 thin anchors, none reached — every one falls below the in-degree ≥6 threshold, which is itself the finding. models/helix-figure (208 words, sources_count: 1, in-degree 4, untouched since 2026-06-06) is the thinnest page in the slice with any reliance on it. Then industries/consulting (214 words, low, 1 source, in-degree 3), industries/accounting (197, low, 1, in-degree 1), industries/construction (152, low, 1, in-degree 0), models/amazon-nova (265, 1, 0), models/magistral (177, 1, 0), models/alphagenome (223, 2, 3), models/genie-3 (325, 2, 2), models/longcat-2 (317, 3, 2). Score 1–2 each.
• The high-reliance, low-word-count band in models/ is a better target than the thin-anchor rule catches. models/deepseek-v3 (608 words, in-degree 21), models/deepseek-r1 (613, in-degree 20), models/kimi-k2 (865, in-degree 20) and models/qwen3 (1,004, in-degree 19) all clear 350 words and so escape the thinness test, while carrying more inbound reliance than anything actioned this run. All four are models/ pages with a model-robustness-check skill built precisely for them. Score 3 each; this is the strongest slice-9 finding and the right first item for the next models/ slice.
• industries/ is under-built relative to models/ by an order of magnitude — 13 pages against 63, with four of the thirteen under 250 words and three at confidence: low on a single source. industries/energy (504 words, in-degree 4) is the one with a live thread behind it after the Lancium and Abbott-interconnection material gathered this run. Score 2.
• companies/source-foundry — scored into the cap and held on the quality gate. Situational Awareness put a further $400M in on August 7, bringing its total to $500M, at a reported $5B valuation, with Stanford founders named. But the company is stealth, has 0 mainspace mentions, and the only account is one WSJ piece behind a paywall. A company page built from a single paywalled report is exactly the thin-and-speculative case the gate exists for. Score 3.
• The AI Tax and Work Protection Act (Casar, Foushee, Jacobs, introduced August 6) — the digest itself flags that its own source was a targeted extraction, with bill number, tax rate and the Work Protection Administration's funding all unestablished. Needs a Congress.gov pull before a legislation/ page is defensible. Score 3.
• DHS — Department of Homeland Security (AI Deployer) expand (385 words, in-degree 17), Defense Innovation Unit (DIU) (483, 13), Five Eyes Joint Guidance on Secure Deployment of AI Agents (May 2026) (511, 12), Taiwan Ministry of Digital Affairs (moda) (586, 11), DOT — Department of Transportation (AI Deployer) (397, 7), NSF — National Science Foundation (AI research funder) (262, low, 1 source, and the NSF 26-513 $100M hubs program landed August 6 with the page still a stub), CBP — Customs and Border Protection (AI Deployer) (199) — all slice-8 carries from 08-07, unreached for a second run. Score 2–3.
• Named in the window with no page: Ike Harris and the Frontier Security Institute (quoted on the August 7 NIST evaluation guidelines, 0 prior mentions); Ryan Calo (quoted on negligence per se under the Washington companion-chatbot law); Yuki Ishizuka and the Washington AG Tech Policy Team; Lucas Hansen and CivAI (second run); Zak Stein and the AI Psychological Research Coalition; Tyler Johnston and the Midas Project (named on two existing pages, still no entity page). Score 1–2 each.
• The NIST AI-evaluation guidelines proposed August 7 and opened for comment — no document number or comment deadline in the retrieved reporting, so not queueable as a source yet. A standards/ candidate once the Federal Register citation exists. Score 3.
• entities/steven-adler (126 words, low, 1 source, untouched since 07-16), entities/susie-wiles (330), entities/sanjog-misra (270) — slice-6 carries, unreached for a fourth run.
• entities/epic, entities/cpsc, entities/national-academies — carried from 08-06. EPIC is cited three times on Youth AI Privacy Act (S. 4199) and authored the model bill behind two legislation/ pages. Score 3–4.
• New person pages implied by the window, none created: clement-delangue, levent-alpoge (both ~7 inbound, both named since 08-05 as matching the in-degree at which prior entity pages were actioned), peter-kirgis, ernie-davis, terence-tao, noam-brown, jasjeet-sekhon, alex-karp, sabrina-ross, charity-clark, tammy-duckworth, lisa-murkowski, plus the ten carried since 08-03.

Standing carries (repeat findings, not per-page gaps)

• The bin/lint-scan.py escaped-pipe target-capture bug — tenth consecutive flag, and it is now actively distorting the health metric. The fix is a .rstrip('\\') on the captured target. It inflates the reported broken-link count by 86 distinct targets (341 reported vs 255 real) and puts pure artifact at the head of the distribution, which is why "broken links" has looked like a standing crisis for ten runs while the real multi-inbound figure is seven. Outside gap-identifier's remit; this scan worked around it locally, as every prior scan has.
• The seven genuinely broken multi-inbound targets, and why none is a gap. concepts/inverse-cooking-problem and concepts/inverse-trust-problem (2 pages each) are curator-blocked coined terms, on needs-review since 07-10 and deliberately not re-actioned. entities/adam-smith and entities/clayton-christensen (2 each) are historical figures cited for framing, not AI actors. entities/ilhan-scheer and entities/samuel-weinbach (2 each) are Aleph Alpha personnel below threshold. wikilinks (4 pages) is a literal-bracket artifact and a lint fix. This list is stable and should be resolved once as a batch rather than re-derived every run.
• An unresolved internal contradiction on Astra, surfaced by today's audit. Open-question 4 says of OpenAI's July 20 long-horizon disclosure that "the company has not said whether the two are related," while Safety and Alignment in an Era of Long-Horizon Models (OpenAI, July 2026) states the model in question is the one announced as having disproved the Erdős unit-distance conjecture — which the Astra page itself ties to the same lineage. Both statements are pre-existing; reconciling them is a lint contradiction job, not an expand.
• New Developments Log/2026-08-06-2212-ai-developments.md is still unprocessed — no matching Operation: Developments-Log entry, second consecutive run to say so. bin/lint-scan.py agrees (unprocessed_devlog=1). Gap-identifier does not fold dev-log files; this run read it for the recent-thread scan only. Its items (the Black Hat Artifactory disclosure, the classified White House oversight framework and its open-model exclusion, the Gemini 3.5 Pro cancellation, the Replit figures, the Weil strict-liability argument) still need the developments-log pass.
• Two schema fits this run had to guess at. companies/lancium is typed company_type: cloud-infra because the enum has no value for a power or data-center infrastructure developer; Wiki/concepts/data-center-siting is used as a tag on that page although data-center-siting is not in the CLAUDE.md tag families (it is already in use on companies/crusoe). Both are curator questions.
• A source-fidelity discrepancy between a digest and its own cited source. The 08-07 22:05 file describes Cassady as "deputy administrator of the NTIA"; its cited source, The Record, says only "a senior official at the National Telecommunications and Information Administration." The new page follows the source. Worth a check on whether the digest enriched beyond its citation.
• The 2026-06-06 cohort reaches models/ and industries/: 11 slice-9 pages combine in-degree ≥2 with last_updated: 2026-06-06, including models/helix-figure, models/alphafold, models/llama-3, models/llama-4, models/alphagenome, models/doubao, models/ernie-4, models/apple-foundation-models, models/genie-3, industries/insurance and industries/construction. Seventh consecutive slice to report this. Migration residue from the v4.3 content/operational split.
• Model-slug inconsistency in models/: claude-opus-45, claude-opus-46, claude-opus-47 beside claude-opus-4-8. Seventh flag.
• companies/doordash placement — fourth run to say the curator ruling on whether prominent AI deployers get company pages is what is missing, not the research.
• Needs-review standing carries: claude-code/claude-cowork placement (07-09); inverse-cooking / inverse-trust coined terms (07-10, curator-blocked); companies/fairly-trained misfile (07-16); the entities/cdao / government/cdao duplicate; models/mai-cyber-1-flash decline (07-28); Andon Labs entity_type (07-30); Situational Awareness LP placement and the entity_type-for-for-profit-non-developer schema gap (08-02).
• The standing source-robustness-check queue, led by Techno-Federalism: How Regulatory Fragmentation Shapes the U.S.-China AI Race (in-degree 89, 1,124 words against a 33,339-word raw file), then California SB 53 — Transparency in Frontier AI Act, Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment), Executive Order 14365 — Ensuring a National Policy Framework for AI, Clawed, California SB 243 — Companion Chatbots, NIST AI Risk Management Framework (AI RMF 1.0).
• [[_meta/briefings/weekly-2026-W19]] cited from mainspace — tenth flag; a schema question for the curator.
• The ingest queue was empty at the start of this run. The one task filed today is the only open item.

Post-run health

bin/lint-scan.py after this scan's edits: banned headers 0, standing Predictions sections 0, header hyperbole 0, wiki self-references 1 (the deliberate index.md lead — one further self-reference was removed from Harvey in passing), dated section headers 28 (unchanged), broken distinct wikilinks 340, down from 341 despite five new pages and three expansions. Zero broken links in any page touched today.

One-line summary

Twenty-four gaps found, nine actioned: five pages created live (NetChoice the highest-in-degree missing page in the wiki at 27 inbound references), three expanded (Astra for the August 7 cyber-capability pause, Legal Services — AI Deployment and Harvey for vendor economics and a valuation contradiction), one foundational source verified and queued for review; an independent audit found and corrected ten fidelity defects, and the slice-9 sweep says the real models/ problem is four heavily-relied-on pages that are too long to trip the thinness rule.