Gap Scan — 2026-07-25

Scanned

Recent window: 4 dev-log files (07-23 08:05, 07-23 22:05, 07-24 09:34, 07-24 22:04 — the last unprocessed, left to tonight's ingest cycle per the run-order precedent carried since 07-22), ~99 pages edited in the last 48h. Rotation slice: 9 — models/ (55 pages) + industries/ (11). Firecrawl unavailable again this run; built-in web search/fetch used with the full authenticity protocol (4 pulls plus 1 search across 6 gaps, well within cap).

The slice and the recent window converged this run. Slice 9's in-degree × thinness scan surfaced four hard thin-anchor flags (models/qwen38-max in-deg 14 at 234 words low/5; models/gemma in-deg 7 at 318w; models/helix-figure in-deg 6 at 208w medium/1 stale since 06-06; models/longcat-2 in-deg 6 at 258w) plus a set of reliance/depth mismatches at higher in-degree (models/deepseek-v3 in-deg 48 at 608w, models/deepseek-r1 in-deg 41 at 613w, industries/defense in-deg 18 at 558w). Independently, the processed 07-24 09:34 dev-log left three type-1 dangling foundational references, two of which are the primary artifacts behind slice-9 model pages that explicitly record having no third-party evaluation. Those three sources and the model pages they anchor were prioritised together.

Gaps actioned (6 of ~21 found)

New pages created (live)

None. Slice-9 pageless candidates resolved to existing slugs after alias resolution; the recent-window pageless candidates (Claude Opus 5, the 25-company "Open Weights and American AI Leadership" letter, the EC TikTok DSA preliminary findings, Anduril's ~$100B round, NJ A-5328) all sit in the unprocessed 07-24 22:04 dev-log and belong to tonight's fold. Per the quality gate, no thin or speculative page was forced.

Pages expanded (live)

• GLM-5.2 (score ~8) — the squarest gap in the slice: the page's Safety and evaluations section read "No third-party (e.g. UK or US AISI) evaluation of GLM-5.2 was available at release," and two government evaluations published the week of the scan supply exactly that. Added a Government cyber evaluations (July 2026) subsection from the UK AISI open-weight cyber-gap post and the UK AISI / CAISI Kimi K3 assessment: AISI's identification of GLM-5.2 as the most cyber-capable open-weight model at time of testing; narrow-cyber-task parity with Opus 4.6 and GPT-5.3-Codex (four months earlier) across all four difficulty levels; cyber-range parity with Opus 4.5 (under seven months earlier), reaching step 7 on fewer tokens than any other model and tracking Opus 4.6 to step 11 before stalling; the 4-to-7-month aggregate lag against the 6-to-10-month 2025 internal baseline; the cost figures ($46 vs ~$85 per 100M-token range run; $6.12 vs $15.17 per reliably solved task); AISI's own elicitation and small-range-set caveats; the CAISI IRT-aggregation methodology PDF; and GLM-5.2's role as the reference point in the Kimi K3 assessment (32% vs 24% on ExploitBench, step 17 vs step 11 on TLO). Added a regulated-by: relationship line and a provenance note. Every prior fact and citation preserved. sources_count 9 → 12; confidence stays medium (the cyber figures sit inside the 3-month fast-decay window and the coding benchmarks remain self-reported). [4 new (Source: URL) cites]
• DeepSeek V4 Pro / V4 Flash (score ~7) — slice thin-anchor at in-degree 27 with 698 words and an Independent evaluation section that stopped at April 2026 commentary. Added a UK AISI cyber evaluation (July 2026) subsection: V4-Pro's narrow-cyber-task parity with Opus 4.5 (five months earlier) and sub-Sonnet 4.5 cyber-range performance; AISI's finding that V4-Pro's safeguards were largely unimpeding, refusing mainly on reverse-engineering tasks and circumvented "simply via a small number of repeat attempts"; and the widest cost-per-capability spread in AISI's comparison ($1.19 vs ~$85 per range run; $0.28 vs $12.50 per reliably solved task), with AISI's non-first-party-hosting and elicitation caveats carried. Repaired frontmatter (safety_case: none-published added, which the model schema requires and the page lacked) and added regulated-by: plus GLM-5.2 and Open-Weight Frontier Models to Relationships. Every prior fact and citation preserved. sources_count 7 → 9; confidence stays medium. [2 new (Source: URL) cites]
• Qwen3.8-Max (score ~5) — the slice's top hard thin anchor: in-degree 14 at 234 words, confidence: low, sources_count: 5. Added a Reported architecture and capabilities section: Shuai Bai's description of Qwen3.8 as the team's first multimodal model above one trillion parameters (text, images, video, documents) and the claimed improvement over Qwen3.7-Max on coding, full-stack development, data analysis and office workflows, both flagged as developer positioning rather than measured results; the sparse-MoE Max-tier lineage (Qwen3-235B-A22B activating 22B per token, Qwen3-30B-A3B roughly 3B) and why the undisclosed active-parameter count, not the 2.4T headline, governs serving cost; the 4-bit ~1.2 TB weights-versus-141 GB-H200 arithmetic and the resulting significance of a smaller-activation, quantised, or distilled variant; and the split developer reception (open-weight competition welcomed, unverified benchmarks and serving feasibility doubted, the timing read against Kimi K3 at 2.8T). Frontmatter parameters corrected to record the sparse-MoE architecture and the undisclosed active count. Every prior fact and citation preserved. sources_count 5 → 7; confidence low → medium (the architecture-class and reception facts are now corroborated across two independent reports, though no benchmark, license, or model card has been published). [2 new (Source: URL) cites]

Link aliases fixed

None. The ≥2-inbound broken targets are unchanged from the 07-20 → 07-24 runs: _meta cross-namespace links (lint-report, log, briefings/weekly-2026-W19), pipe-escape table artifacts (companies/anthropic\, companies/openai\, gpt-53-codex\, claude-mythos-5\), literal illustrative text inside provenance notes (sources/..., wikilinks), already-queued sources with explicit provenance notes (sources/claude-sonnet-5-system-card — wired to INGEST-claude-sonnet-5-system-card-2026-06-30.md; sources/trustworthy-agents-in-practice), and standing carries (claude-code, claude-cowork, in-degree-2 person pages). Alias resolution was run against all folders on concepts/iterative-deployment, concepts/ai-history, entities/federal-reserve, entities/ecb, and entities/bank-of-england: none resolves to an existing page under another name, so each is a genuine missing-page candidate rather than an alias error, and each carries in-degree 2 — below the action threshold. They stay in the deferred backlog.

Queued — foundational sources

• UK AI Security Institute, "How Far Behind the Frontier are Leading Open Weight Models on Cyber?" (July 17, 2026) (score ~9) — type-1 dangling foundational reference; the "4 to 7 month" figure is named on Open-Weight Frontier Models, UK AI Safety Institute (AI Security Institute), and US-China AI Competition: Different Races, Different Metrics with supporting cites only, and is now load-bearing in the federal open-weight-restriction debate. Verified: aisi.gov.uk/blog (the institute's own domain, DSIT footer present), full text fetched; the 4-to-7-month conclusion, the 6-to-10-month 2025 baseline, the TLO range description, the cost table, and the Kimi K3 forward-look confirmed verbatim; outbound NCSC and AISI links resolve; corroborated by the NIST/CAISI Kimi K3 assessment, which hyperlinks this URL and cites its central finding. Date correction recorded: the 07-24 09:34 dev-log carried this item under a July 24 date; the canonical page is dated July 17, 2026. Saved: Raw Sources/How Far Behind the Frontier are Leading Open Weight Models on Cyber - AISI.md. Queued: INGEST-aisi-open-weight-cyber-gap-2026-07-25.md.
• UK AISI / CAISI, "Preliminary Assessment of Kimi K3's Cyber Capabilities" (NIST, July 23, 2026) (score ~9) — type-1 dangling foundational reference on the window's most-cited thread and the first joint UK AISI / CAISI publication the wiki tracks; folded onto Kimi K3 and Open-Weight Frontier Models by the 07-24 cycle with supporting cites, also named on GPT-5.6 (Sol, Terra, Luna), Index, and Project Glasswing: An initial update. Verified: nist.gov (official U.S. government host; CAISI sits within NIST/Commerce), full text fetched; Drupal metadata gives published 2026-07-23T08:00-04:00 and "Released July 23, 2026"; the step-17-of-32 vs 28.5 result, the 0/41 vs 20/41 ACE finding, the safeguards finding, and "Solves of TLO are no longer exclusive to a small set of models" confirmed verbatim; ExploitBench resolves to arXiv 2605.14153 (CMU) and the CAISI GLM-5.2 methodology PDF resolves on the NIST document server. Methodological caveats recorded in the verification trail (preliminary; U.S. closed-weight models evaluated with system-level safeguards disabled; K3's aggregate score from a single benchmark). Saved: Raw Sources/UK AISI CAISI Preliminary Assessment of Kimi K3 Cyber Capabilities.md. Queued: INGEST-aisi-caisi-kimi-k3-cyber-2026-07-25.md.
• OpenAI, "Safety and alignment in an era of long-horizon models" (July 20, 2026) (score ~8) — type-1 dangling foundational reference; the 07-22 cycle folded "long-horizon post detail" onto OpenAI with a supporting cite, and the episode is named on Overview, AI Autonomy Risk, and AI for Science. It is the developer's own account of the sandbox-pause episode upstream of the evaluation-security cluster already queued, and the artifact the Kill Switch Act debate reaches back to. Verified: openai.com/index/ (canonical URL and og:url match), full text plus both footnotes fetched; the "PowerCool" cooldown, the hour-long sandbox-vulnerability search behind public PR #287, the token-fragmentation example, and the redeployment findings confirmed verbatim; modded-nanogpt PRs 287/291/294/300/305/307/309 are public GitHub identifiers and the Erdős announcement resolves on the same domain. Classification note carried: this is a developer self-disclosure, so its severity characterisations are attributed positions, to be set alongside the independent readings (Zvi; Transformer). Saved: Raw Sources/Safety and Alignment in an Era of Long-Horizon Models - OpenAI.md. Queued: INGEST-openai-long-horizon-safety-2026-07-25.md.

Authenticity-verification failures

None. All three pulled sources verified on canonical primary hosts. One metadata conflict was resolved rather than rejected: the AISI open-weight cyber-gap post's date differs between the 07-24 dev-log (July 24) and the canonical page (July 17); the canonical date governs and the correction is recorded in both the INGEST task and the verification trail.

The supplementary research for Qwen3.8-Max surfaced a large volume of SEO/content-farm reproductions of the Qwen3.8 announcement (aggregators publishing "benchmarks" for a model whose benchmark table has not been released). None was used. The two sources folded are the MarkTechPost report, which distinguishes explicitly between what Alibaba confirmed and what it claimed and quotes the Qwen developer's own X post, and SCMP's release coverage; both are cited as supporting sources. Alibaba's own announcement remains a set of X posts with no model card, license, or benchmark table, so no primary artifact exists to queue.

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

• Unprocessed 07-24 22:04 dev-log — tonight's fold first, then gap-scan candidates. Claude Opus 5 is the highest-value item: Anthropic released Claude Opus 5 on July 24 at $5/$25 per million tokens with a per-request effort setting, reporting more than double Opus 4.8 on Frontier-Bench v0.1, within 0.5% of Fable 5's peak CursorBench 3.2 at half the cost per task, 3× the next-best model on ARC-AGI 3, a 2.3 automated-behavioral-audit score, and cyber classifiers expected to intervene ~85% less often than Fable 5's with fallback to Opus 4.8. No wiki page names it yet; it becomes a score ~8 new-page candidate the moment tonight's cycle folds it, and the anthropic.com release post a source-queue candidate. Also pending: the 25-company "Open Weights and American AI Leadership" letter (a substantive coalition letter — score ~7 source-queue candidate once folded, and directly relevant to the AISI evaluations queued today); the EC TikTok DSA preliminary findings; the Digital Omnibus on AI reaching Official Journal publication; Anduril's ~$100B round; Warner's six-bill AI package; NJ A-5328 data-broker law (legislation page candidate ~4); the House FY27 NDAA 216–212 passage; the Google DeepMind / Isomorphic Labs biosecurity position.
• Slice-9 residue (three-per-run pacing): DeepSeek-V3 (in-deg 48 at 608w — the largest reliance/depth mismatch left in the slice, ~5), DeepSeek-R1 (in-deg 41 at 613w, ~5), Qwen3 (in-deg 40, high/5 at 1,004w, ~4), Kimi K2 (in-deg 39, medium/6 at 866w, ~4), Defense / Military — AI Deployment (in-deg 18 at 558w, last_updated 07-13, with an outdated Anduril row — "$800M+ cash burn" — that tonight's fold will make stale-by-contrast, ~5), Gemma (Google open-weight models) (in-deg 7 at 318w, ~3), Helix (Figure AI Vision-Language-Action model) (in-deg 6 at 208w, medium/1, stale since 06-06, ~3), LongCat-2.0 (Meituan) (in-deg 6 at 258w, ~3), IsoDDE (Isomorphic Labs) (in-deg 12 at 446w, ~3), Retail — AI Deployment (in-deg 12 at 633w, stale 06-12, ~3), Energy and Electric Power Sector (504w, medium/4, stale 06-22, on a very live thread but only in-degree 3, ~3), Construction — AI Deployment (152w, low/1, in-deg 1, ~2).
• industries/manufacturing — topical hole (gap type 6, score ~3). Manufacturing-as-deployer appears substantively across Defense / Military — AI Deployment (the GDOTS–Deterrence defense-industrial-base agreement), Chinese AI Policy (the 14th Five-Year Plan's 70%-AI+Manufacturing-by-2027 targets), and the embodied-AI thread, but no industries/ page anchors it and no [[industries/manufacturing]] link exists to signal demand. Deferred rather than actioned: with zero inbound links the demand signal is inferred from prose alone, and the material currently sits legibly on the pages that carry it. Revisit if the embodied-AI/physical-automation thread produces a third or fourth page's worth of manufacturing-specific content. Lower-signal siblings at in-degree 2: industries/software, industries/consulting, industries/automotive, industries/agriculture, industries/accounting.
• Genuine missing-page candidates confirmed by alias resolution but below the in-degree threshold (2 each): concepts/iterative-deployment (would gain a strong anchor from the OpenAI long-horizon source queued today — flagged in that INGEST task's touch points), concepts/ai-history, concepts/safety-training-methodologies, concepts/economic-possibilities-for-artificial-intelligence, concepts/ai-coarse-grainings, concepts/cross-national-ai-policy-tracking, concepts/algorithmic-decisionmaking, entities/federal-reserve / entities/ecb / entities/bank-of-england (all three named from International Monetary Fund (IMF)).
• Carried from prior runs: slice-8 residue (Character.AI Litigation (umbrella) high on sources_count: 0 at in-deg 9; DHS — Department of Homeland Security (AI Deployer); Defense Innovation Unit (DIU); frontmatter misratings on ICE — Immigration and Customs Enforcement (AI Deployer), OECD Recommendation on Artificial Intelligence, DARPA (Defense Advanced Research Projects Agency)); slice-7 legislation thin anchors (California SB 53, New York RAISE Act, Executive Order 14110 (RESCINDED), EU General-Purpose AI Code of Practice (2025), Bletchley Declaration (AI Safety Summit, 2023), TAKE IT DOWN Act); slice-6 entities O–Z thin anchors; H.R. 9619 (People-First Chatbot Act) primary text (congress.gov empty body ×4 — not retried this run); METR "expenditure horizon", Zitron "Subprime Data Center Crisis", Pethokoukis transformative-AI essay (essay/research queue candidates ~4); the Kill Switch Act congress.gov bill number (check at ingest); Substack-redirect URL upgrades; the orphan-check sources-layer backlog (six state chatbot-act primary texts plus the Model Spec blog post).
• Standing carries (needs-review): claude-code / claude-cowork placement (curator-blocked, 07-09); inverse-cooking / inverse-trust coined terms (07-10); companies/fairly-trained misfile (07-16); entities/cdao / government/cdao duplicate (lint's lane); person/org pages at in-degree 2.

One-line summary

Slice 9's audit and the week's evaluation-security thread converged: queued the three verified primary artifacts of the government cyber-evaluation cluster (the UK AISI open-weight cyber-gap post — with a date correction against the dev-log — the joint UK AISI / CAISI Kimi K3 assessment, and OpenAI's long-horizon safety disclosure), and used the first two to close the explicit "no third-party evaluation" hole on GLM-5.2 and the stale independent-evaluation section on DeepSeek V4 Pro / V4 Flash, while lifting the slice's thinnest high-traffic model anchor, Qwen3.8-Max, from 234 words and low confidence to a sourced account of what Alibaba has and has not actually published; Claude Opus 5 has no page and is the top deferred item pending tonight's fold of the unprocessed evening dev-log.