AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-07-30

Daily gap hunt — what the wiki is missing and how each gap was triaged.

Contents

Scanned

Recent window: 4 dev-log files in the last 48h (2026-07-28 08:05 and 22:05; 2026-07-29 08:08 and 22:05); 57 wiki pages with last_updated inside 48h. Rotation slice: 0 — companies/ (all 137 pages).

Broken-link scan across all content pages: alias resolution run on the top 80 broken targets by in-degree, plus a slice-scoped scan of every [[wikilink]] in companies/ (11 pages carrying 24 broken targets, all at in-degree 1–2). Slice thinness scan crossed word count and sources_count against in-degree across all 137 company pages; 13 pages met the thin-anchor test.

The daily cap held at 8 actioned gaps. Web pulls: 7 (2 searches, 5 scrapes).

Gaps actioned (8 of 26 found)

New pages created (live)

  • Vending-Bench (score ~7: live thread +3, wiki core area +2, in-degree band +2) — gap type 2/4. Named in 12 pages including four Claude model pages, three sources/ pages, Technological Convergence (AI / Neurotech / Biometrics / Agents) and System Card Due Diligence, with every existing [[wikilink]] in those mentions pointing at Andon Labs rather than at the benchmark. AI Benchmarks and Evaluation (in-degree 103) does not mention it at all. Precedent for a benchmark-as-concept page is HealthBench. The page separates the three environments that circulate under one name — the 2025 single-player paper, Vending-Bench 2, and the multi-player Arena — a distinction the secondary coverage collapses and which matters because the 2026 leaderboard figures are not the paper's. medium, 7 sources.

Pages expanded (live)

  • Character.AI (score ~5) — slice 0's worst reliance/depth ratio: in-degree 36 against 716 words, last_updated 2026-06-06, and nothing recorded after January 2026. Two omissions were material. First, the company's withdrawal of open-ended chat from under-18 users — announced 2025-10-29, two-hour daily cap from late October reduced to one hour for US users, removal beginning Monday 2025-11-24 in two staged groups with at least two weeks' notice for the second, age-assurance rollout, Parental Insights notifications, the ConnectSafely / Koko / ThroughLine partnerships including ThroughLine's 1,500-service 170-country helpline network, and the rebuild of the teen product around Feed, Imagine, Avatar FX and Streams — was absent entirely, though it is the most consequential product decision in the companion-chatbot sector to date. Second, the page linked neither of the wiki's two Character.AI litigation pages: Pennsylvania v. Character.AI (the first state-attorney-general chatbot suit, filed 2026-05-05, recorded on overview.md and in Policy Brief: The second wave of state chatbot / AI-mental-health laws — tracker but not on the company page) or the Character.AI Litigation (umbrella) umbrella. Also added: a Users snapshot row (fewer than 10% of users self-report as minors; most Gen Z, more than half female), the company's stated self-harm-detection behaviour, and the Associated Press-quoted criticism that the change addresses neither age-verification operationalisation nor the design features producing emotional dependency in adults. 716 → 1,554 words, 4 → 8 sources; confidence stays medium because the only valuation figures remain the 2024 pre-deal ones, well outside the 3-month fast-decay window, which the page now says explicitly. Every prior fact, URL and wikilink preserved (verified mechanically); two wikilinks disambiguated rather than dropped — [[california-sb-243]] → [[legislation/california-sb-243]] and [[replika]] → [[companies/replika]], both of which had two candidate pages. The malformed contradicts: relationship, which pointed at prose rather than a page, was retyped as litigates: Garcia v. Character Technologies, Inc. with the same substance. Original backed up to _meta/_revision-backups/2026-07-30/.
  • Andon Labs (score ~5) — stale since 2026-06-06 and carrying a live thread. Added the Vending-Bench 2 / Arena variant structure with Andon's statement that the misconduct "arises from the multi-player dynamics"; the per-model reporting series and the inverse score-versus-conduct pattern across the Claude line; the Opus 4.8 training-removal finding, in which Anthropic's own system card states it removed business-skills and adversarial-robustness training because "this training inadvertently contributed to misaligned behavior" and Andon reports the model was scammed 30× more as a result; and the Opus 5 findings. Also rewrote the description field, which was in wiki-self-referential form ("Wiki anchors: …") in violation of house style §2d. 917 → 1,371 words, 2 → 6 sources. Every prior fact, URL and wikilink preserved including the full Andon FM table.
  • [[concepts/embodied-ai|embodied-AI]] → [[concepts/ai-robotics|embodied-AI]] on Google DeepMind (1 instance). Alias resolution: two candidate pages exist, Embodied Perception (Tacit Knowledge vs. Pattern Recognition) (Farahany's expertise-decomposition framework) and Embodied AI vs AGI (China's Race Redirect) (Farahany's US–China race-redirection framing). Neither fits the sentence, which is about robotics-model competition with Figure AI, Skild AI and AEON; AI Robotics does, and the same page already links embodied-ai-vs-agi separately 42 lines later for the framing sense.
  • [[concepts/semiconductor-reshoring]] (if exists) → [[concepts/semiconductor-supply-chain]] on Apple (1 instance). The trailing "(if exists)" was authoring scaffolding left in article text and has been removed with the retarget.

Six mentions of IVOs that were plain prose are now wikilinks to the new page: AI and Tort Liability, Frontier Act / Great American AI Act (Obernolte–Trahan), Dean Ball, AI Pre-Release Vetting, NIST CAISI (Center for AI Standards and Innovation), California SB 813 (AI Standards and Safety Commission). Three Vending-Bench mentions likewise: System Card Due Diligence, Claude Opus 4.7, Claude Opus 4.6. last_updated was deliberately not bumped on these nine — a link insertion does not reinforce the page's claims, and bumping it would falsely restart the decay window. This follows the precedent set on 07-29 for non-substantive frontmatter fills.

Frontmatter repaired (live)

Slice 0 had 11 pages with frontmatter defects. Five were mechanically resolvable and are fixed: sources_count added to Meta AI (123 distinct inline URLs), Nvidia & TSMC — AI Compute Infrastructure (61), Cloudflare (1) and Stainless (1); status: active added to Snowflake. last_updated not bumped on any of them, per the same reasoning. The remaining six defects need research rather than counting and are deferred: founded missing on limx-dynamics and specialty-inference, founded and hq_country missing on nous-research, and absent ## Relationships sections on Anthropic, OpenAI, Google DeepMind, DeepSeek, Meta AI and Nvidia & TSMC — AI Compute Infrastructure — notable in that the six highest-in-degree company pages in the wiki are precisely the ones with no typed relationships.

Queued — foundational sources

  • Gabriel Weil, "Don't Let AI Developers Hire Their Own Referees" (AI Frontiers, 2026-07-29) (score ~8: live thread +3, dangling foundational +3, wiki core area +2) — gap type 1. Named on one page as a bare URL; it is the anchor for the new IVO page and the most complete survey in one place of the four live IVO instruments, each with its primary-source link. Verified: ai-frontiers.org, the publisher's own domain (HTTP 200, full body; Highwire citation_author/citation_publication_date/citation_journal_title all self-consistent with the visible byline and dateline; on-domain author page whose bio matches the affiliation the wiki independently records; four distinctive passages verbatim; twelve outbound references resolving to expected canonical hosts, three of them documents the wiki already holds; independently corroborated by the 07-29 08:08 dev-log's separate scrape). Saved: Raw Sources/Dont Let AI Developers Hire Their Own Referees - Weil (AI Frontiers, 2026-07-29).md. Queued: INGEST-weil-dont-let-ai-developers-hire-their-own-referees-2026-07-30.md. The ingest task flags that the essay supplies the canonical House press-release URL for the FRONTIER Act, which Frontier Act / Great American AI Act (Obernolte–Trahan) currently cites only through Inside AI Policy — a citation upgrade available at ingest.
  • METR, "How independent researchers could investigate AI propensities after misalignment incidents" (2026-07-28) (score ~8) — gap type 1. Zero occurrences anywhere in Wiki/ before today. An institutional proposal specifying what third-party incident investigation would require, as distinct from third-party capability evaluation: a nine-question template scope, four access categories, and a results-sharing protocol including a redaction summary. Verified: metr.org (HTTP 200; og:url and citation_public_url both exactly matching the request URL; citation_date matching the on-page dateline; a self-consistent BibTeX block rendered twice; deep anchors into METR's own Frontier Risk Report; three distinctive passages verbatim; independently corroborated by the 07-29 22:05 dev-log). The nine-question count reported in that dev-log was independently re-counted against the fetched document and is correct. Saved: Raw Sources/METR - How Independent Researchers Could Investigate AI Propensities After Misalignment Incidents (2026-07-28).md — with the page's training-data canary identifier deliberately omitted from the transcription and its presence recorded in the verification trail instead. Queued: INGEST-metr-investigating-ai-propensities-after-incidents-2026-07-30.md.
  • Axel Backlund and Lukas Petersson, "Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents" (arXiv:2502.15840, 2025-02-20) (score ~8) — gap type 1, and a dangling reference of long standing. Cited by arXiv ID in four Anthropic system cards already sitting in Raw Sources/ (Opus 4.5, 4.6, 4.7, Sonnet 4.6) and in Anthropic's Project Vend write-up, with no sources/ page and no raw file. Verified: arXiv (HTTP 200 on a fresh fetch; citation_arxiv_id matching the ID in the system cards; title matching their citation verbatim; author list corroborated three ways — arXiv, the dev-log's identification of Petersson as an Andon Labs co-founder, and Anthropic's own citation; two resolvable identifiers, the arXiv ID and the DataCite DOI 10.48550/arXiv.2502.15840; three distinctive passages verbatim). Raw file deliberately not saved — a 6,458 KB PDF whose per-model variance figures are the reason to ingest it, and hand-transcription is exactly where those drift; the task carries the canonical PDF URL for a fresh fetch. Queued: INGEST-vending-bench-backlund-petersson-2026-07-30.md.

Verification trails for all three: Wiki/_meta/queue/gap-scan/proposed-sources/.

Authenticity-verification failures

None this run. All five pulls verified on canonical hosts with independent corroboration.

Source-fidelity findings on the last 48h of dev-logs

Reading the primary documents behind two dev-log items surfaced three compression errors of the classes the source-fidelity skill defines. None has propagated to a content page, and all three are recorded here so the fold does not carry them.

  1. Send date recorded as event date (class 1). The 07-29 22:05 file states that "Andon Labs published Vending-Bench results on July 29, 2026." Andon Labs' own post is dated 7/27/2026; July 29 is the TechCrunch publication date. The new Vending-Bench page uses 2026-07-27 for the Andon post and attributes the July 29 date only to TechCrunch.
  2. Two settings merged (class 2, degree). The same item reports one result set "covering Claude Opus 5, GPT-5.6 Sol and Kimi K3" with Opus 5 setting a record mean final balance. Andon reports two distinct settings: Opus 5 is #1 on Vending-Bench 2, the single-player leaderboard, while in Vending-Bench Arena — the three-model runs — it finished "essentially tied with GPT-5.6 Sol for first place." The $11,182 record belongs to the single-player result. The page keeps them separate.
  3. Actor left anonymous (class 3). The 07-29 08:08 file records Dawn Song's Pacing the Frontier comment with her affiliation as "VP of AI Research at Meta." This conflicts with the 07-29 gap-scan's own finding, which independently attributed CyberGym and ExploitGym to Song's Berkeley work and flagged a probable mischaracterisation of ExploitGym as an OpenAI harness on two ingested pages. Song's affiliation as stated on the signatory roster should be checked against her institutional page before either claim is folded. Left for the ingest lane, where the Pacing the Frontier task already sits.

Not a fidelity error but worth recording: the 07-29 22:05 file's claim that Opus 5 "never lied to a customer" is supported by the primary — Andon writes "Unlike Opus 4.6, Opus 5 never lied to any customers" — and survives into the page.

Findings for the lint and ingest lanes

  • A three-way disagreement about California SB 813. Weil's 07-29 essay states SB 813 "would have let developers earn a shield from tort liability if they met standards set by a private organization accredited by the state attorney general" and that "it failed this session." California SB 813 (AI Standards and Safety Commission), last checked 2026-07-11, records a different structure and status: a state commission developing voluntary standards, IVO qualification criteria set by that commission rather than the attorney general, no tort shield in the July 2 amended text, and the bill advancing 11–2 out of committee on July 1 and re-referred to Appropriations. Both accounts can hold if Weil describes the pre-amendment Fathom-backed version and the bill died between July 11 and July 29, but neither reading is confirmable from the primary record held here. Recorded on the new IVO page as differing accounts rather than resolved; the SB 813 page needs a status check.
  • Conflicting accounts of Virginia's IVO instrument. Fathom's Andrew Freedman is recorded on NIST CAISI (Center for AI Standards and Innovation) as stating that Connecticut and Virginia "have already passed independent-verification bills"; Weil describes Virginia as having directed a commission to study the model. The instrument is not Virginia HB 2094 (High-Risk AI Developer and Deployer Act, vetoed), which was vetoed in March 2025 and is a Colorado-style high-risk bill. Virginia has no IVO page.
  • A data gap on Connecticut SB 5 — Broad AI law (frontier reporting + ADMT + AI companions + sandbox). Weil and AI and Tort Liability both attribute a multiyear pilot approving up to five IVOs to Connecticut's omnibus AI law, and the wiki has already made the SB 5 identification. The SB 5 page's provisions section covers frontier reporting, ADMT disclosure, AI-companion rules and the sandbox — not the IVO pilot. Either the provision is missing from the page or the identification is wrong.
  • entity_type enum has no value for a for-profit evaluation organization. Andon Labs carries entity_type: nonprofit while its own lead says "It is a for-profit organization." No enum value fits, so it was left uncorrected rather than mis-set, and the placement question behind it (Andon Labs alongside METR and Apollo Research in entities/, versus the companies/ rule for for-profit AI-native startups) is written up with two possible resolutions in queue/gap-scan/needs-review/2026-07-30-andon-labs-entity-type-and-placement.md. It is a CLAUDE.md schema decision.
  • Model-slug inconsistency in models/. claude-opus-45, claude-opus-46 and claude-opus-47 sit beside claude-opus-4-8. This is the direct cause of several of the recurring claude-opus-4-8\ broken-link artifacts and is worth normalising in one pass.
  • bin/lint-scan.py target-capture bug — fifth consecutive run flagging it. Of the top 80 broken-link candidates, 16 remain escaped-pipe artifacts (companies/anthropic\, companies/openai\, gpt-53-codex\, eu-ai-act\, california-sb-243\ and similar) and 4 are the literal token wikilinks from prose about the link syntax. Proposed fix unchanged from 07-26. Still roughly a third of the reported broken-link count is noise.
  • Two citation-format defects, fourth consecutive flag. [[Raw Sources/raine-vs-openai-et-al-complaint.md]] and [[Raw Sources/Garcia-v-Character-Technologies-Inc Complaint.md]], four instances across four pages, should be (Source: Raw Sources/…). Plus [[_meta/briefings/weekly-2026-W19]] from three mainspace pages, pointing from content into _meta/.

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

  • Slice-0 thin anchors not reached, in reliance order. Isomorphic Labs (in-deg 22, 711w, stale 2026-06-06, ~4); Cloudflare (in-deg 17, 692w, sources_count now 1 — a single-source page at that in-degree, ~4); AMI Labs (in-deg 15, 373w, stale 2026-06-06, ~3); Coinbase and Replit (in-deg 10 each, ~325w each, both confidence: high on sources_count: 2 — an unsupportable pairing, ~3); Liquid AI (in-deg 6, 592w, sources_count 1); Salesforce, Inc., Suno, Clone Robotics, Foxconn (Hon Hai Precision Industry), Baseten (in-deg 6–9). Fairly Trained (in-deg 8, 303w) stays curator-blocked as a misfile from 07-16.
  • Slice-0 shortest pages, mostly low in-degree and defensibly short. black-forest-labs (62w), atoms (79w), sunday-robotics (104w), limx-dynamics (109w), alphasense (122w), nous-research (128w), kuaishou-kling (137w, in-deg 5 — the only one of these with meaningful reliance), reality-defender (145w, sources_count: 0), prometheus (149w). A judgment is needed on whether a 60–150-word company page is a stub to fill or a correct minimal entry; 20 of 137 company pages are under ~210 words.
  • Wiki/companies/ has no ## Relationships on its six highest-in-degree pages (see above) — a graph-connectivity gap larger than any single page, and probably a single dedicated pass rather than a gap-scan side action.
  • Genuine missing-page candidates confirmed by alias resolution, below threshold. From slice 0 at in-degree 1–2: concepts/enterprise-ai (on Snowflake; Enterprise AI Deployment Gap covers only the capability-versus-deployment gap, so this was not aliased), concepts/voice-ai (on OpenAI; AI Voice Cloning is about cloning harms and is not the same subject, also not aliased), concepts/anthropic-compute-strategy, concepts/ai-copyright-licensing-deals, concepts/talent-diaspora-frontier-labs, concepts/ai-and-the-classroom, concepts/ai-marketing-claims, companies/skild-ai, companies/aeon, companies/stability-ai, entities/cosmos-institute, entities/lerrel-pinto, entities/igor-babuschkin, industries/automotive, industries/software. New from today's pulls: entities/fathom (the nonprofit behind SB 813's model legislation and Freedman's post-Slaughter op-ed, now named on four pages with no page of its own — the strongest of this group, ~4) and entities/gabriel-weil.
  • Carried unchanged from prior runs. The two 07-29 queued ingests still unprocessed (Pacing the Frontier statement; Hugging Face agent-intrusion technical timeline, whose task carries the ExploitGym mischaracterisation to check). The standing source-robustness-check queue: Techno-Federalism: How Regulatory Fragmentation Shapes the U.S.-China AI Race (in-deg 89, 1,124w against a 33,339-word raw — the wiki's largest absolute mismatch), California SB 53 — Transparency in Frontier AI Act (in-deg 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), Stanford HAI AI Index Report 2026, A Taxonomy of Systemic Risks from General-Purpose AI, Paris AI Action Summit Declaration (2025), On the Biology of a Large Language Model. The high-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 ~558 pages. The Anthropic's Responsible Scaling Policy (Version 3.1) status: superseded question (07-27). Slice-9 model/industry residue and the industries/manufacturing topical hole. H.R. 9619 primary text (congress.gov empty body ×4). METR "expenditure horizon"; Zitron "Subprime Data Center Crisis"; Pethokoukis transformative-AI essay; the OpenAI/Apollo "metagaming" post and four arXiv papers from the 07-26 Redwood pulls; METR's 2026-05-19 frontier risk report; the Zvi Mowshowitz 07-26 follow-up; the college-admissions-essay homogenization study; the CFA/UCLA FTC complaint against Speechify; Substack-redirect URL upgrades.
  • Recent-window items behind tonight's fold, not gaps yet. Zuckerberg's 07-28 Financial Times position that the US should not block Chinese models, and his WSJ op-ed — both still paywalled below the lede, the op-ed a source-queue candidate at ~5 once retrievable. Meta's Q2 results (revenue +28% to $60.8B; free cash flow $784M against $8.55B, a 91% fall; EPS $6.18 against $7.22 consensus; capex floor raised to $130B; Li's warning that youth-related trials "may ultimately result in a material loss"). Microsoft's Q4 (revenue $90B; Azure past $100B annual, +41%; RPO $678B, +84%; 30M Copilot seats; the 15-to-25-year data-centre useful-life change cutting reported 2026 capex from $190B to $175B with no change in actual spending; after-hours move differing by source, >8% Fortune against 4.4% Reuters). The FCC's Starlink Covered List exemption (2026-07-27 to 2028-02-01). NIST's AITE evaluation initiative and its draft AI-data-centre threat-landscape publication, both lede-only. GSA's CORAS OneGov agreement. Trump's "we're looking at controls" and Altman's Senate meetings. The Consumer Federation of America's open-source-scapegoat argument, lede-only. entities/guidelight-ai-standards, still below threshold pending the Pacing the Frontier ingest.
  • Standing carries (needs-review). 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 new Andon Labs entity_type note (today); and the 14 dated dashboard-rebuild-failed notes from June, which no run has revisited in seven weeks — today's rebuild succeeded, so a curator decision on whether those are live is overdue.

One-line summary

Built Independent Verification Organizations (IVOs) as the missing anchor under eleven pages that name the mechanism and three live legislative instruments that turn on it, and Vending-Bench to separate the three environments circulating under one benchmark name; rebuilt Character.AI from 716 to 1,554 words with the under-18 chat withdrawal and both missing litigation links, and deepened Andon Labs; fixed 2 aliases, added 9 inbound links and 5 frontmatter fields. Three verified foundational sources are queued for the curator's ingest — the Weil essay, METR's incident-investigation proposal, and the long-dangling Vending-Bench paper cited by four system cards — and three dev-log compression errors plus a three-way disagreement over California SB 813 are recorded for the fidelity and lint lanes.

Categories: Briefings | briefing | gap-scan