Gap Scan — 2026-08-01

Scanned

Recent window: 4 New Developments Log/ files covering 2026-07-30 and 2026-07-31, and 92 wiki pages edited in the last 48 hours. Rotation slice: 2 — concepts/ G–O (43 pages).

Broken-link analysis ran over 1,710 content pages and 359 raw broken targets. Applying the trailing-backslash strip that four consecutive scans have recommended to bin/lint-scan.py collapses that to 279 distinct targets and leaves only 9 with an in-degree of 2 or more — confirming again that roughly 22% of the reported broken-link count is the Obsidian table pipe-escape artifact, not a wiki defect. All nine survivors were already on the deferred backlog or curator-blocked, so no genuine link-alias error surfaced this run and no alias fix was made.

Gaps actioned (8 of 31 found)

New pages created (live)

• ATOMIC Act (H.R. 9965) — H.R. 9965, the AI Threat Output and Monitoring Incident Containment Act, introduced 2026-07-27 and recorded in the 07-31 evening digest with no page behind it. Score ~5 (live thread +3, US AI regulation core area +2). Built from Congress.gov and the sponsors' joint announcement. Provenance note recorded: the full bill text was not yet posted to Congress.gov, so provision-level statements rest on sponsor descriptions and contemporaneous reporting and should be upgraded once the primary text is available. confidence: medium.
• CHAT Act and CHAT Act 2.0 — covers both the September 2025 CHAT Act (S.2714) and the tiered CHAT Act 2.0 introduced 2026-07-29. Score ~5. The wiki already held People-First Chatbot Act, Kids Internet and Digital Safety Act (KIDS Act, H.R. 7757) and California SB 243 — Companion Chatbots but had no page for the federal bill those measures are most often compared against. Provisions of S.2714 come from the Congress.gov bill text; the 2.0 provisions from the sponsor's announcement. confidence: medium.
• Gemini Robotics 2 — the 2026-07-30 Google DeepMind release. Score ~4. The nightly fold recorded it inside AI Robotics and Google DeepMind, but a three-model family with published per-task success rates, a new safety benchmark and a separate safety technical report had no models/ page. Built entirely from the DeepMind announcement, including the developer's own stated limits ("multi-finger dexterous manipulation remains challenging"; movement speed). confidence: low, sources_count: 1 — single-source, and the safety technical report, model cards and ASIMOV-Agentic dataset the announcement names were not retrieved, which the page states in its lead.
• Inkling-Small — the 2026-07-30 Thinking Machines Lab open-weights release. Score ~4. Inkling existed and was updated by the nightly fold, but the point release carries its own architecture, pricing and a full published eval suite. Three benchmark tables transcribed from the primary announcement, with the company's two harness caveats preserved and all figures attributed as developer-reported. confidence: low, sources_count: 1.

Pages expanded (live)

• Government AI Procurement — the highest-reliance page in the rotation slice (in-degree 25) and stale since 2026-06-06. Expanded 874 → 1,513 words; sources_count 6 → 12. Added: the GSA–xAI OneGov agreement at $0.42 per organization (2025-09-25) with its FedRAMP and DoD Impact Level upgrade path, which also supplies the missing context for the page's existing WIRED item about xAI losing a government contract six weeks earlier; a new authorization-pathway subsection covering GSA's Buy AI guidance and FedRAMP's AI prioritization criteria; a new Procurement rulemaking section on the March 2026 proposed GSAR "American AI Systems" clause and its seven contractor obligations; the DIU seven-to-three portfolio narrowing; and the DOD supply-chain risk designation and Judge Lin's remark under Risks and criticism. 8 new (Source: URL) citations. The ## Relationships block gained a deployed-by: line linking nine government/ pages, and three "planned" markers were cleared for pages that now exist. Every pre-existing fact and citation preserved; original backed up to _meta/_revision-backups/concepts/. confidence held at high — not raised, and defensible on the 6-month policy decay window now that it has been reinforced.

Link aliases fixed

None. See the scan note above — after backslash normalization no unambiguous alias error remained.

Queued — foundational sources

• "SOTA alignment assessments don't strongly update us against misalignment" (Alexa Pan, Redwood Research, 2026-07-31) — score ~8, the run's highest: live thread (+3), dangling foundational reference the wiki names by name (+3), safety/alignment core area (+2). The 07-31 evening digest retrieved it as a page summary only, so the wiki records the argument's existence without holding it. Verified: blog.redwoodresearch.org (HTTP 200; title, byline and 2026-07-31 timestamp corroborated by page metadata and by independent citation in the digest; the distinctive Anthropic quotation and the post's adoption of Anthropic's term "reliability of the assessment" both confirmed present). No raw file saved — the markdown fetch exceeded the retrieval budget, so per the skill's large-document provision the task carries the verified canonical URL for a fresh fetch at ingest. Verification trail: Wiki/_meta/queue/gap-scan/proposed-sources/redwood-pan-sota-alignment-assessments-2026.md. Queued: INGEST-redwood-pan-sota-alignment-assessments.md.

Source-fidelity corrections made against the digest

Two claims in the 2026-07-31 evening digest did not survive contact with the primary sources, and the new pages carry the corrected versions:

• Actor left incomplete. The digest attributed H.R. 9965 to Rep. Celeste Maloy alone. Maloy is the sponsor of record on Congress.gov, but the bill is bipartisan and was announced jointly with Rep. Sara Jacobs (D-CA-51) as co-lead. The digest also omitted the bill's short title, the ATOMIC Act, under which both sponsors' offices refer to it.
• Enforcement understated. The digest recorded the CHAT Act 2.0 as "enforceable by the Federal Trade Commission." The sponsors' announcement states enforcement authority for the FTC and state attorneys general. Both readings are recorded on the page with their sources, since the narrower one comes from the same trade-press article the digest used.

Neither is a digest failure of the kind source-fidelity was written for — both trace to writing from a single trade-press restatement rather than the sponsor and Congress.gov primaries — but both are the compression-and-anonymization pattern that skill names, appearing four days after it was added to the schema.

Authenticity-verification failures

None. Five documents were pulled this run (Congress.gov H.R.9965, Congress.gov S.2714 text, husted.senate.gov, sarajacobs.house.gov, deepmind.google, thinkingmachines.ai, blog.redwoodresearch.org) and all resolved on their canonical primary hosts at HTTP 200 with corroborating metadata. No aggregator or mirror was substituted.

One partial-retrieval caveat is recorded rather than hidden: the Redwood post's full body could not be captured, and a structured extraction was used for verification only. That extraction returned paraphrase rather than the post's sentences, and at least one apparent quotation looks like a position the post is characterizing rather than asserting. The INGEST- task carries an explicit instruction not to fold anything from the verification record into a page.

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

• Slice-2 thin anchors not reached, in reliance order. Levels of Autonomy for AI Agents (Feng-McDonald-Zhang framework) (in-deg 12, 856w, sources_count: 1, stale 2026-06-06, ~4 — a well-built page whose only defect is single-sourcing, so expanding it needs reception/uptake sources rather than more of the same primary, and it was cut to keep one expansion substantive); Normative Competence (in-deg 12, 1,149w, sc 1, stale 06-06, ~2); Model Welfare (in-deg 8, 515w, the slice's only confidence: low, ~3); Mathematical Impossibility of Perfect Fairness (in-deg 9, 689w, sc 1, ~3); The Long Industrial Revolution (in-deg 7, 482w, sc 1, stale 06-06); Metagaming (in-deg 7, 611w, sc 1 — pairs naturally with the un-pulled OpenAI/Apollo metagaming post already on this list).
• The 2026-06-06 cohort, now measured on a second consecutive concepts/ slice. Eight slice-2 pages combine in-degree ≥10 with last_updated: 2026-06-06, the date of the v4.3 content/operational split: Government AI Procurement (expanded today and off this list), Liar's Dividend (18), Model Context Protocol (MCP) (15), LLM Fallacy (14), Jagged Frontier (13), Levels of Autonomy for AI Agents (Feng-McDonald-Zhang framework) (12), Normative Competence (12), Mixture of Experts (MoE) (10). Slice 1 reported the same pattern at 24 pages on 07-31. Two slices in, this is a folder-wide migration residue rather than a set of individual gaps, and is better handled as a planned pass than as daily gap-scan residue — the second run to say so.
• Recent-window model and legislation candidates below the cap. models/deepseek-v4-flash (V4-Flash moved to public beta 07-31 with nine published agent-benchmark scores; DeepSeek V4 Pro / V4 Flash is the landing page, ~3); models/minimax-h3 (video model, 07-31; MiniMax holds it, ~3); Nano Banana 2 and its 48-hour withdrawal from Google Earth after the BBC Verify fabricated-imagery report (a AI and Misinformation fold or a models/ page, ~3); Epoch AI's FrontierMath: Open Problems expansion (page retrieved as summary only, so the problem count is unknown — a research pull before a page, ~2); the National Quantum Readiness Act and the reintroduced AI Ads Act, both recorded on ATOMIC Act (H.R. 9965) as related but pageless.
• Foundational-source candidates below the cap. Gary Marcus's 07-31 response to Anthropic's incident report, with the collected Bill Gurley and Zack Korman reactions (op-ed, borderline under the heuristic, ~4); Alex Kantrowitz's 07-31 argument that labs may withhold their most capable models from metered APIs (retrieved as summary only, ~3).
• Genuine missing-page candidates confirmed by alias resolution but below the in-degree threshold, carried forward unchanged from 07-31. Concepts: economic-possibilities-for-artificial-intelligence, algorithmic-decisionmaking, ai-and-language-models, talent-flow-china-us, data-broker-regulation, a-vision-of-democratic-ai, sortition, safety-training-methodologies, ai-coarse-grainings, cross-national-ai-policy-tracking, ai-history, biometric-identification, enterprise-ai, voice-ai, and inverse-cooking-problem / inverse-trust-problem (curator-blocked as coined terms since 07-10). Person pages: anil-seth, orin-kerr, eric-goldman, kevin-klyman, lee-anne-fennell, michael-j-d-vermeer, sacha-altay, joseph-bernstein, roge-karma, alondra-nelson, clayton-christensen, adam-smith, dina-powell-mccormick, raja-krishnamoorthi, samuel-weinbach, ilhan-scheer, david-luan, girish-gupta, alex-mallen, caleb-biddulph, dan-lahav, andrew-freedman, bri-treece. New from today's pulls: celeste-maloy, sara-jacobs, jon-husted, andy-kim — four legislators now named on pages created today, none of whom has an entity page, and entities/ai-policy-network, the ATOMIC Act's only recorded endorser. Institutions: federal-reserve, ecb, bank-of-england, carnegie-mellon, santa-fe-institute, sais, uw.
• [[_meta/briefings/weekly-2026-W19]] referenced from three mainspace pages (AI Macro-Prudential Policy, Nathan Lambert, Index). Sixth consecutive flag, still a schema question for the curator rather than a mechanical rewrite: whether mainspace may cite a briefing at all.

Carried unchanged from prior runs

• Queued ingests still unprocessed: the "Pacing the Frontier" employee statement and the Hugging Face agent-intrusion technical timeline (07-29, the latter carrying an ExploitGym mischaracterisation to check); Weil, METR and Vending-Bench (07-30); the five 07-31 queue items (claude-1-launch, first-genai-copyright-suits-2023, garante-chatgpt-suspension-2023, google-bard-launch, llama-1-paper). Today's Redwood task makes twelve open.
• 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 (234), Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment) (178), Executive Order 14365 — Ensuring a National Policy Framework for AI (144), Clawed (105), California SB 243 — Companion Chatbots (98), NIST AI Risk Management Framework (AI RMF 1.0) (74).
• The high-confidence-on-sources_count: 1 question, measured at 57–65% across three consecutive sources/ slices and still awaiting the one-time curator policy call that would settle roughly 558 pages.
• The bin/lint-scan.py target-capture bug — fifth consecutive flag, and quantified again above at 80 of 359 targets. The one-line fix remains \[\[([^\]|#]+?)(?:\\?\|[^\]]*)?\]\], or a trailing-backslash strip on the captured target.
• Needs-review standing carries: claude-code / claude-cowork placement (07-09); inverse-cooking / inverse-trust coined terms (07-10); companies/fairly-trained misfile (07-16); entities/cdao / government/cdao duplicate; models/mai-cyber-1-flash quality-gate decline (07-28); the Andon Labs entity_type and placement question (07-30), which today's models/ additions do not bear on. Model-slug inconsistency in models/ (claude-opus-45/46/47 beside claude-opus-4-8), third flag.
• The 14 dated dashboard-rebuild-failed notes from June, unrevisited for seven weeks. The 07-31 rebuild succeeded, as did today's; the notes can be cleared on a curator's word.

Post-run verification pass

An independent audit of all five touched files against house-style, source-fidelity and CLAUDE.md was run before this report was filed. It found real defects; all were repaired. Findings acted on:

• Confidence and source counts were overstated on both new models/ pages. Gemini Robotics 2 carried confidence: medium / sources_count: 3 and Inkling-Small medium / 2, when each rests on exactly one vendor announcement. Both corrected to low / 1 per the gap-identifier rule that a page built from a single source takes low. This was the run's most consequential defect: it misrepresented the evidentiary base under fast-decay benchmark claims. The safety_case field on the Gemini page also asserted a safety technical report that had been named but never retrieved; it now says so, and both pages carry an explicit single-source disclosure in the lead. sources_count on the two legislation/ pages was likewise corrected 4 → 3 to match the distinct sources actually cited.
• A benchmark contradiction between Inkling-Small and Inkling. The existing Inkling page records Inkling-Small figures — Terminal Bench 2.1 52.7%, IFBench 83.4%, HLE-with-tools 46.6% — that differ from the new page's 64.7%, 82.2% and 47.8%. Both are correct: the older set are preview-checkpoint figures from the July 15 release, the newer the released model. Neither page said so. Both now label them.
• Four misused typed relationships. regulated-by was inverted on ATOMIC Act (H.R. 9965) (asserting the bill is regulated by OpenAI, Anthropic and Google) and misapplied on CHAT Act and CHAT Act 2.0; supersedes was used as free prose asserting its own negation on CHAT Act and CHAT Act 2.0 and Inkling-Small; deploys-in on Gemini Robotics 2 pointed at two concept pages rather than a deployer; and the new deployed-by: line on Government AI Procurement pointed the wrong way for a concept page. All recast as related or instance-of with the intended meaning stated inline.
• Five uncited factual passages — the CHAT Act 2.0 prohibition list, the AI Policy Network endorsement, the March 2026 DOD designation and suit dates, the Inkling-Small open-weights comparison paragraph, and a claim that Gemini Robotics 2 is built on the Gemini 3 / Gemini 3 Pro generation. Four were given their sources; the Gemini-3 claim was not supported by any source and was removed.
• A contradiction with California SB 53. The ATOMIC Act page set its investment threshold against "the revenue thresholds used in SB 53," but SB 53's primary trigger is a 10^26 FLOP compute threshold, with the $500M revenue test applying only to its strictest tier — so the sentence set SB 53 against a category it belongs to. Rewritten and cited.
• Editorial and self-referential residue. A bolded maintainer-directive block on the ATOMIC Act page ("when this page was written… should be upgraded") rewritten as neutral prose; "at the time of writing" removed from three pages; "GSA nonetheless signed the OneGov agreement" — an implied criticism neither source makes — replaced with the plain date; "the principal drafting change" reattributed to the sponsor's office; a Texas Character.AI incident hedged as an allegation; and an off-topic paragraph on the National Quantum Readiness Act and the AI Ads Act, carried over from the digest's round-up source, deleted.
• Three paywalled insideaipolicy.com citations now carry the subscriber-wall flag that source-fidelity requires and that the sibling Anthropic v. United States (Pentagon ban challenge) page already used; the unnamed July 28 summit is now marked as unnamed in the reporting. Executive Order 14319 — Preventing Woke AI in the Federal Government is now wikilinked where the new text referenced it in bare form.

Revision fidelity on Government AI Procurement passed cleanly — the diff against _meta/_revision-backups/ is purely additive, with zero facts, citations or sections dropped, and the three "planned" markers promoted into live links rather than deleted.

Two findings were noted and not acted on, as pre-existing rather than introduced by this run: five tags on the procurement page (procurement, gsa, onegov, usai, federal-ai-adoption) sit outside the CLAUDE.md tag families, and its region/actor tagging predates the run. Left for lint.

One-line summary

Four new pages and one substantial expansion committed live, one foundational Redwood Research source verified and queued for review, and two source-fidelity corrections applied against the 07-31 digest — an incomplete sponsor attribution on H.R. 9965 and an understated enforcement clause on the CHAT Act 2.0 — with the concepts/ 2026-06-06 migration cohort and the lint-scan.py regex bug both flagged for the second and fifth consecutive run.