Gap Scan — 2026-08-15

Scanned

Recent window: 3 New Developments Log/ files inside 48 hours (2026-08-13 08:12 and 22:11; 2026-08-14 22:12) and 45 content pages carrying last_updated: 2026-08-14 or later.

Rotation slice: 2 — concepts/ G–O, 43 pages, scanned for in-degree-weighted thinness, confidence-versus-sourcing mismatch, and citation density.

Broken-link analysis: 249 distinct broken targets across 277 instances. Twenty-eight carry an in-degree of 2; none carries 3 or more. All twenty-eight went through alias resolution.

Queue state at start of run: empty apart from README.md and the excluded subdirectories — the 2026-08-14 queue-processor pass drained both open tasks.

Run-order note: New Developments Log/2026-08-14-2205-ai-developments.md was unprocessed at the start of this scan. The 2026-08-14 Developments-Log operation covered files through 2026-08-13-2205. Four of the seven gaps actioned below — the GLM-5.3 page, the Anthropic risk report, the CyberGym benchmark's August figures, and the Pax Silica exclusivity draft — draw on that unfolded file, which is why they scored as untouched live threads. The nightly fold should still run over it and may find these items already handled here.

Gaps actioned (7 of 34 found)

New pages created (live)

• CyberGym — score 8 (dangling foundational reference the wiki names by name +3, live thread +3, core area +2). The highest-in-degree dangling reference currently in the graph: CyberGym is named by name on twelve content pages — models/glm-5, models/gpt-55, models/claude-opus-46, models/claude-sonnet-46, models/gpt-56, models/muse-spark, and the Mythos, Opus 4.6, Opus 4.7, Opus 4.8, Muse Spark and Pacing-the-Frontier sources/ pages — with no concepts/ page and no Raw Sources/ file behind it. It is the reference score in every cross-lab cyber comparison the wiki carries, including the Mythos High Cybersecurity threshold trigger. Precedent for a benchmark page in concepts/ is established by Vending-Bench and concepts/healthbench. Built from the arXiv abstract, the Berkeley RDI project page and cybergym.io, plus the twelve existing wiki pages that report scores against it. confidence: medium. Depends on queued source: CyberGym (arXiv:2506.02548) — carries a > **Provenance note:** recording that its inline citations should upgrade on ingest.

• GLM-5.3 — score 5 (live thread +3, core area +2). Launched August 14, 2026 and absent from models/, where GLM-5, GLM-5.2 and GLM-4 all have pages. Built from Z.ai's own announcement as the primary source, corroborated by Reuters, Times of India and MarkTechPost. confidence: medium; open_weights: pending, since the weights had not been released. The page records the delayed-weight-release decision, Z.ai's disclosed evaluation settings, and that no independent evaluator had reproduced the figures.

Pages expanded (live)

• Model Welfare — slice-2 thin anchor, score 5 (in-degree 9 → +2, core safety area +2, confidence: low with sources_count: 2). Was 515 words with no coverage of the deprecation-and-preservation commitments that are the concept's most concrete instance, and no critique of the assessment methodology. Added ## Welfare assessment as a system-card category (Haiku 4.5, Opus 4.7 and Mythos cards, cited to existing wiki pages), ## Deprecation, retirement and preservation (Anthropic's published commitments; the Claude 3 Opus retirement), Mowshowitz's methodological objection to lab-run welfare assessments, and ## Open questions. 3 new (Source: URL) cites plus 7 internal wikilinks. confidence low → medium; sources_count 2 → 7. See the duplication finding below — this page overlaps concepts/ai-welfare, and the expansion was deliberately confined to the applied/methodological side.

• Pax Silica (US-Led AI Supply-Chain Initiative) — score 5 (live thread +3, core area +2). Added ### Exclusivity covering the State Department's drafted letter to the 35 signatories of the June 2026 "AI Opportunity Statement," stating that Pax Silica membership "cannot be held alongside membership in duplicative initiatives whose expectations conflict with our own," and the Kazakhstan dual-membership detail. Modality preserved: the letter is drafted, undated and not yet sent. Also repaired frontmatter — the page was missing sources_count entirely. 1 new cite; sources_count added at 7.

• Zhipu AI, GLM-5.2, Autonomous cyber-agents — graph wiring for the two new pages: a GLM-5.3 product entry on the company page with the cyber figures and the release-delay decision, a superseded-by relation and a CyberGym related on GLM-5.2, and a depends-on: [[concepts/cybergym]] on the autonomous-cyber-agents page, whose cross-lab capability figures all come from that benchmark. last_updated bumped on all three; zhipu-ai sources_count 9 → 10.

Link aliases fixed

None this run. The 28 multi-inbound broken targets were all resolved as either genuine but low-priority person/institution gaps (see Deferred) or as targets with no existing page under any name — no alias errors were found.

Queued — foundational sources

• Zhun Wang, Tianneng Shi, Jingxuan He, Matthew Cai, Jialin Zhang and Dawn Song, "CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale" — named by name across twelve pages with no source page. Verified: arXiv (arxiv.org), corroborated against the Berkeley RDI project page and cybergym.io; identifier arXiv:2506.02548, ICLR 2026 OpenReview 2YvbLQEdYt. Raw file: not saved — PDF fetch deferred to ingest to stay inside the daily web-pull cap. Verification record: Wiki/_meta/queue/gap-scan/proposed-sources/cybergym-wang-song-2026.md. Queued: INGEST-cybergym-wang-song-2026.md.

• Anthropic, "Redacted Risk Report August 2026" — the primary text behind the "very low" → "low" misalignment-risk change carried in the 2026-08-14 22:05 digest. Verified: anthropic.com, the developer's own domain; full document retrieved (~449,000 characters) with its numbered-claim structure confirmed; corroborated by Axios, 2026-08-14. Raw file: not saved — document length exceeded what this scan should carry inline; the canonical URL is confirmed reachable and complete. Verification record: Wiki/_meta/queue/gap-scan/proposed-sources/anthropic-risk-report-aug-2026.md. Queued: INGEST-anthropic-risk-report-aug-2026.md.

Authenticity-verification failures

None. Both sources pulled this run passed all six protocol steps. Two discrepancies were flagged rather than resolved, and neither is grounds for rejection:

• CyberGym zero-day counts disagree between the authors' own two primary hosts. The arXiv abstract states "34 zero-day vulnerabilities and 18 historically incomplete patches"; the Berkeley RDI project page states "35 zero-day vulnerabilities and 17 historically incomplete patches." Both are author-controlled and neither explains the difference. CyberGym reports both figures and attributes each to its host; the ingest should check the ICLR camera-ready.
• Z.ai's August 2026 benchmark table contradicts its own prose. The GLM-5.3 announcement's prose attributes the 83.8% CyberGym figure to Mythos 5, and Reuters reported it that way, while the table in the same announcement heads that column "Fable 5 (w/ fallback)." Anthropic has published no Mythos 5 CyberGym figure that would settle it. Both CyberGym and GLM-5.3 record the discrepancy rather than picking a reading.

Needs review

• concepts/model-welfare and concepts/ai-welfare are near-duplicate pages. Found while researching the model-welfare expansion. Both open by defining "model welfare," both anchor to Anthropic's programme and the Emotion Concepts paper, neither links to the other, and Wiki/index.md lists only ai-welfare. A defensible split exists (moral patienthood versus applied methodology), so this is a curation call and merging is supersede's job. The expansion stayed on the applied side and added a scope note plus a cross-link; nothing was moved or deleted from either page. Full note with a suggested resolution: Wiki/_meta/queue/gap-scan/needs-review/2026-08-15-model-welfare-ai-welfare-duplication.md.

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

• The escaped-pipe scanner artifact — confirmed closed, third and final recording. A first-pass scan this run reported 335 broken targets across 395 instances; 86 of those, at 92 instances, were [[folder/slug\|Alias]] links whose escaped pipe the naive \[\[([^\]\|#]+) capture reads as a trailing backslash. Every one resolves to an existing page. The corrected scanner reports 249 targets across 277 instances, which is the figure used above. This matches the finding the 2026-08-13 and 2026-08-14 runs recorded; it is a scanner artifact, not a wiki defect, and any future run reporting ~335 targets is running an uncorrected scanner.

• Slice-2 thin anchors not reached — 4 of the 5 the slice surfaced (score 3–4 each). concepts/normative-competence (in-degree 10, confidence: high on sources_count: 1 — a confidence-versus-sourcing mismatch, the strongest of the four), concepts/mathematical-impossibility-perfect-fairness (in-degree 7, high/1 — same mismatch), concepts/long-industrial-revolution (in-degree 7, 482 words, medium/1), concepts/metagaming (in-degree 6, 611 words, medium/1). Two of these carry confidence: high on a single source, which the decay table does not support at any window; that is a lint finding as much as a gap-scan one.

• concepts/algorithmic-decisionmaking (score 3, rising for a fourth consecutive run). In-degree 2 as a broken target from sources/hai-policy-brief-ai-health-insurance. Carried since 2026-08-12.

• companies/riot-platforms (score 3, fourth consecutive run). Carried unchanged.

• Compute-as-collateral and GPU rental futures (score 3, fourth consecutive run). Carried unchanged. The 2026-08-14 Nvidia/OpenAI Ohio financing restructure — the guarantee reduced from $250 billion to under $120 billion for half the 10-gigawatt build-out — is a live instance and strengthens the case.

• Author and institution pages implied by sources/ pages, none created (score 2 each). The 28 multi-inbound broken targets are almost entirely of this kind: entities/dawn-song and a Berkeley RDI page (now implied twice over by CyberGym), entities/kevin-klyman, entities/orin-kerr, entities/alondra-nelson, entities/eric-goldman, entities/lee-anne-fennell, entities/michael-j-d-vermeer, entities/sacha-altay, entities/joseph-bernstein, entities/clayton-christensen, entities/samuel-weinbach, entities/ilhan-scheer. This is a standing structural pattern — the wiki ingests a source and links its authors without creating them — and has now been recorded across five consecutive runs without being addressed. It is better handled as one batch pass than six per run.

• Institution shorthand targets (score 1–2). entities/carnegie-mellon, entities/santa-fe-institute, entities/sais, entities/uw, entities/federal-reserve, entities/ecb, entities/bank-of-england. Several are abbreviations (sais, uw) that would be better rewritten to full names than created as pages.

• Concept targets named but pageless (score 2 each). concepts/inverse-cooking-problem and concepts/inverse-trust-problem — carried since 2026-08-09 and flagged in needs-review/2026-07-10-inverse-cooking-trust-coined-terms.md as coined terms, which is why they have not been created. concepts/a-vision-of-democratic-ai, concepts/data-broker-regulation, concepts/talent-flow-china-us, concepts/economic-possibilities-for-artificial-intelligence, concepts/ai-and-language-models.

• Named in the 48-hour window with no page (score 1–3 each). Legislation: the GATE Act (Harrigan, Aug 14), the China AI Power Report Act (S. 5382, Husted, Aug 7), the Military Readiness Through Resilient Lands Act (H.R. 10069, Aug 6) — all three from a single Nextgov round-up, none yet with independent coverage. Models: OpenAI's Ultrafast mode for GPT-5.6 Sol (Aug 13), Gemini 3.7 Flash (Aug 13, already folded per the 2026-08-14 log). People: Sarah Friar, Dali Rajic, Denise Dresser, Amelia Glaese, Dylan Scandinaro, Jim Banks, Christopher Phelan, Gabriel Wagner, Ara Kharazian. Concepts: ExploitBench and ExploitGym, currently covered inside CyberGym rather than separately, which is the right placement until either accumulates independent coverage.

• Anthropic's February 2026 risk report (score 3). The baseline against which the August report's rating change is measured; the wiki holds neither. Flagged inside the August report's ingest task as a companion candidate rather than queued separately this run, to stay inside the cap.

• Recent-window items that are developments-log work, not page gaps. The Anthropic 2028 revenue projection ($190–200bn) and the reported $2tn listing expectation; the OpenAI enterprise-over-consumer revenue crossover and $40bn annualized run rate; Apple's China-market model trained with Alibaba's support; the Taiwan MoDA AI-assisted intrusion disclosure (government/taiwan-moda already updated); the Flock retention and audit-tool changes (companies/flock-safety already updated); the Databricks close (already folded). Each belongs to the nightly fold, and most already ran.

Self-audit of this run

• One wiki self-reference introduced into mainspace and removed. The first draft of GLM-5.3 carried "a policy question the wiki tracks under Open-Weight Frontier Models." bin/lint-scan.py reported self_refs=2 against a pre-existing baseline of 1; the clause was rewritten to "the recurring policy question of open-weight frontier release" and the re-run returned self_refs=1. The remaining instance is pre-existing and not from this run.
• Broken-target count reconciliation. This report uses 249 distinct broken targets from the scan built for this run; bin/lint-scan.py reports 250. The one-target difference comes from scope — the lint scanner includes index.md and overview.md, which the gap scanner treats as catalog pages. Neither figure moved as a result of this run's edits: the two new pages resolved no previously-broken target, because neither had been linked before it existed.
• Dashboard rebuilt. python3 bin/build-dashboard.py completed, writing 1,926 article pages; concepts/cybergym.html and models/glm-5-3.html confirmed present in the output.

One-line summary

Two pages created — CyberGym, the wiki's highest-in-degree dangling reference, and GLM-5.3 — five expanded, two foundational sources verified and queued for review, and one duplication finding (model-welfare / ai-welfare) left for the curator to resolve.