Spokes.wiki Search About

Log — AI Governance Wiki

Append-only history. Entries start with ## [YYYY-MM-DD] <op> | <title>.

[2026-06-26] ingest | Akrites — industry OSS-security body (hub-routed)

Hub-routed Telegram drop (msg 599, thenewstack.io). WebFetch returned only page chrome twice → firecrawl fallback (HUB edge rule) to get the article body past the cookie banner. T3 (reputable trade press + Linux Foundation press-release backstop; one Endor stat unverified). Strong-fit ingest: the wiki already held the Fable 5/Mythos ban and the "controls don't stop proliferation" critique that Akrites directly answers. Created akrites (Organization), akrites-launch-tns (NewsArticle source), and a thin linux-foundation venue node (the operational counterpart to ietf). Synthesis: added "governance by collective industry defense" as a new instrument — a second non-state venue (defense-side/operational vs IETF's input-side rule-writing), the industry counterpart to the state's export ban on the same incident, and the constructive corollary of the efficacy critique (harden targets vs gate the tool). Cross-linked the export-ban + efficacy pages. +3 pages.

[2026-06-07] split | ai-governance-wiki created from _inbox (3 sources)

Spun out of the hub ai-governance _inbox cluster once it reached 3 cohering sources (per ../HUB.md). Founding sources (all Telegram-dropped 2026-06-07): nemko-ai-governance-services (vendor AI-GRC + certification, aligned to EU AI Act / ISO 42001 / NIST RMF), israel-ai-regulation-overview (Israel's sectoral soft-law regulatory framework), and israel-ai-strategy-2026 (Israel Innovation Authority's national AI strategy, "Excellence Outweighs Scale," April 2026 draft). Scaffolded raw/ + wiki/, CLAUDE.md from the hub template, and these spine files; added the registry block to ../wikis.md. The strategy PDF is held in raw/AI-Strategy-2026-ENG.pdf (the other two are URL-only). Founding page set: 3 source summaries + concept pages (ai-governance, eu-ai-act, iso-iec-42001, nist-ai-rmf, risk-based-regulation, responsible-innovation) + orgs (nemko-digital, israel-innovation-authority).

[2026-06-07] ingest | founding sources (Nemko, Israel regulation, Israel AI strategy)

Ingested the three cluster sources per Ingest. See synthesis for the thesis: AI governance spans an organizational axis (management-system standards, assurance, certification — Nemko) and a national axis (regulatory regimes + AI strategy — Israel), with a live tension between the EU's horizontal binding law and Israel's sectoral soft-law model. Israel's two sources show the regulation/strategy pair: a risk-based, OECD-aligned, sandbox-driven approach that deliberately avoids an EU-style horizontal AI act to stay "export-ready."

[2026-06-07] ingest | LoFrayer vs. Israel Bar Association (AI traffic-fine startup)

Hub-routed (Telegram, Calcalist/Ctech). New source lofrayer-bar-association (NewsArticle, url): the Israel Bar Association's 72-hour ultimatum to shut LoFrayer (AI fine-appeal startup) for "unauthorized practice of law" under Section 20 of the 1961 Bar Association Law. New concept professional-licensing-and-ai (DefinedTerm) — a third governance vector: existing professional- licensing law as de facto AI regulation (distinct from new AI acts and national strategy), and the public-protection-vs-access/anti-monopoly tension. Updated israel-ai-regulation-overview (concrete instance of "existing statutes reinterpreted for AI"), ai-governance (new mechanism), and synthesis (new "regulation by existing law" section + new tension). Index: +1 concept, +1 source. Caveat: single news report of an unadjudicated 72-hour ultimatum.

[2026-06-09] ingest | +2 international standards (OECD AI Principles, CoE Framework Convention) — all-spokes cron test

Filled the "whose standards win?" open question with the international layer the spoke lacked: oecd-ai-principles (DefinedTerm, src — the first intergovernmental AI standard; the shared AI-system definition the EU Act / CoE treaty / US / UN all reuse; voluntary soft law) and framework-convention-on-ai (Legislation, src — Council of Europe; the first legally binding international treaty on AI, signed by EU+US+UK; principles-based, treaty not regulation). Reframed the binding tier as two shapes (product-regulation vs state-treaty) above the voluntary OECD/ISO/NIST tier. (UNESCO Ethics Recommendation attempted but its Wikipedia URL 404'd — deferred, not padded.) 13 → 15 pages.

[2026-06-10] ingest | China + US AI regulation — all-spokes pass (the fork becomes a 4-way map)

Two national regimes filling the "whose standards win / national models" gap. china-ai-regulation (DefinedTerm, source, Wikipedia Regulation-of-AI) — state-driven, content-binding: CAC algorithm registration (1,400+ filed by 2024), the 2023 Generative AI Interim Measures (one of the first national GenAI frameworks), and 2025 mandatory AI-content labeling/watermarking (the hard-law end of the provenance spectrum). us-ai-policy (DefinedTerm, source, same) — market-driven/fragmented: Biden 2023 safety-testing EO rescinded, Trump Dec-2025 EO preempting state laws (California/Colorado), no federal horizontal law. Together they turn the synthesis's horizontal-vs-sectoral fork into a four-way instrument-philosophy map (EU rights-driven horizontal-binding · China state-driven content-binding · US market-driven fragmented · Israel innovation-driven sectoral-soft) and add a new federal-vs-subnational preemption axis (US). Convergence caveat holds — all lean on the oecd-ai-principles definitional substrate. Folded into synthesis (new 2026-06-10 section) + index (2 DefinedTerm rows). No contradictions. 15 → 17 pages.

[2026-06-12] ingest | UK AI regulation — "pro-innovation approach" (GOV.UK White Paper)

All-spokes daily expansion. Added uk-ai-regulation (@type DefinedTerm) — the fifth jurisdiction on the instrument-philosophy map (EU/China/US/Israel + UK). The UK is pro-innovation, cross-sector top-down: five central principles (safety-security-robustness; transparency/explainability; fairness; accountability/governance; contestability/redress) delegated to existing sector regulators, no single AI law. Key analytic payoff: shows "sectoral-soft" was two mechanics — UK top-down delegated principles vs Israel bottom-up statutory reinterpretation. Sharpens the "does soft law hold under pressure?" question (AI Safety→Security Institute rename Feb 2025; statutory frontier framework mooted 2026 = soft→binding drift). Convergence caveat intact (principles track oecd-ai-principles). Wired to eu-ai-act / responsible-innovation / us-ai-policy / china-ai-regulation; synthesis "four-way → five-way" note; index updated. 1 new page. Authoritative (UK government White Paper).

[2026-06-14] ingest | US export-control ban on Anthropic models (Fable 5 / Mythos 5) — TechCrunch

Routed from hub (Telegram drop). Reported US export-control crackdown (12 Jun 2026) on two Anthropic models, triggered by an Amazon-CEO→Treasury cyber-misuse report + a David Sacks "jailbreak" framing; Anthropic counters the capability "already exists in other public models." Quality gate: tier T4 (trade press, heavily "reportedly"/secondhand — justified: breaking event, no primary/official source yet); soft-gate ingested with the contested-ness flagged. Not previously paged → new source summary anthropic-export-ban-2026 + new concept export-controls-on-ai (export control as a governance instrument — the lever beyond conduct rules). Gap-relevance: extends the instrument taxonomy and complicates the us-ai-policy "deregulatory" cell; integrated into us-ai-policy + synthesis (new instrument note + a "deregulatory yet interventionist" contradiction). Cross-spoke: the regulated models/providers (Anthropic/Amazon/Fable5/Mythos5) noted as ../llm-providers-wiki territory, not duplicated. freshness: volatile. url provenance. Site rebuild + commit follow.

[2026-06-18] ingest | The Free and Open Web Is Under Attack at the IETF (EFF)

Telegram-dropped; hub-routed here (runner-up: ../search-marketing-wiki — robots.txt/crawler access is SEO-adjacent, but the substance is AI policy, not search marketing). Source summary eff-web-under-attack-ietf (T2, advocacy POV; url provenance). New instrument: governance by technical standard — the ietf as a non-state venue writing AI access rules in protocol form, targeting AI's inputs (training-data access) rather than conduct or distribution. Concept pages ai-preferences (robots.txt-evolution "don't train" signals) + web-bot-auth (cryptographic bot ID with selective-access double-use). Gap-relevance: extends the instrument taxonomy with a first non-state venue + a first input-side lever; sharpens "Whose standards win?" into a new open question about who governs the open web. Integrated into synthesis (instruments-beyond-conduct-rules bullet, lineage 06-18, new open question) + ai-governance hub page. Entities (2/8 budget): electronic-frontier-foundation (NGO, publisher) + tori-noble (Person, author) — both new; IETF paged as subject Thing, not a discovered agent. Cross-spoke: output-side content labeling sits in china-ai-regulation; robots.txt/crawling overlaps ../search-marketing-wiki (noted, not duplicated). Site rebuild + commit follow.

[2026-06-18] ingest | DefinedTerm enrichment pass (subagent)

Thinnest-first deepening of DefinedTerm pages with fetched, authoritative sources. Sources fetched: NIST AI RMF official landing page (nist.gov, T1) + NIST AIRC trustworthiness characteristics page → new source summary nist-ai-rmf-overview; A-LIGN ISO 42001 explainer (a-lign.com, T2) → new source summary iso-42001-align-explainer; oecd.ai/en/ai-principles re-fetched (T1, already a source page). Enriched: nist-ai-rmf — added AI RMF 1.0 release date (2023-01-26), the four-function detail (Govern as cross-cutting; Map→Measure→Manage), the seven trustworthiness characteristics, and the GenAI Profile NIST-AI-600-1 (2024-07-26); stub note removed. iso-iec-42001 — added Annex A–D structure, the six PDCA requirement areas, and the not-yet-harmonized-vs-EU-AI-Act status; stub note removed. oecd-ai-principles — added the verbatim OECD AI-system definition and the May-2024 update rationale, explaining the definitional-substrate thesis. risk-based-regulation — sharpened the NIST entry and added an OECD-definition entry, both sourced. responsible-innovation — Related footer + OECD cross-link, no body padding (already developed). Index: new TechArticle source group with the two new pages. All [[links]] verified against existing slugs. Ran avoid-ai-writing over new prose. No build, no commit (orchestrator handles).

[2026-06-21] ingest | "From PGP to Mythos" — export-control efficacy critique (hub-routed)

Telegram drop (hub msg 546, TechCrunch). Adds the efficacy dimension export-controls-on-ai flagged as open: a 3-era history (PGP/Crypto Wars → spyware/Wassenaar → Mythos/AI) arguing cyber export controls don't stop proliferation (relocation, non-signatory havens, enforcement discretion, capability replication; rare win = FinFisher 2022). New source cyber-export-control-history (TechArticle, T2 — thesis-driven; historical cases solid, the "it'll fail again" prediction is the author's). Deepened export-controls-on-ai with a "does it even work?" section + a 2nd open thread (governance theatre vs. buys-time); cross-linked anthropic-export-ban-2026 (the Mythos ban it critiques). Refresh-in-place, no new concept page. +1 page.

[2026-06-28] ingest | US restricts OpenAI's GPT-5.6 access (The New Stack) — hub-routed

The export-control thread's next beat. Two new source/Thing pages + three updates + two new entity nodes.

  • openai-gpt56-access-restriction (NewsArticle, T3, thenewstack.io) — White House limits GPT-5.6 to government-approved partners, approved customer-by-customer during preview (CNN + The Information
    • Altman staff Q&A); the 2nd lab hit two weeks after Anthropic.
  • eo-14409 (Legislation) — the instrument the Anthropic story lacked a name for: Executive Order 14409 (2 Jun 2026), 30-day pre-release submission of frontier models. Reframes the Anthropic ban as the first enforcement beat of a standing pre-release-review regime, not a one-off.
  • Entities: sam-altman (Person, OpenAI CEO — regulated-party voice) and the-new-stack (NewsMediaOrganization — recurring publisher; owner Insight Partners invests in OpenAI+Anthropic, disclosed conflict). Reused cross-wiki: openai, glm-52, david-sacks.
  • Updated us-ai-policy (security-lever exception → named-EO regime, 2 beats), export-controls-on-ai (Open Thread #1 "is it repeatable?" → partly resolved: yes, under EO 14409), anthropic-export-ban-2026 (new "instrument behind it" section), and synthesis (the security-lever bullet rewritten around the named, repeated regime + the efficacy/open-weight-accelerant thread).
  • Cross-spoke (../llm-providers-wiki): access-denial on closed frontier models is a policy accelerant for open-weight/Chinese challengers (GLM-5.2 beats GPT-5.5 on coding, self-hostable) — noted as context, not duplicated. Caveat recorded: per-model application is secondhand; EO text is the verifiable kernel.

[2026-06-28] ingest | GPT-5.6 Preview system card (OpenAI) — hub-routed

The primary source the GPT-5.6 thread was missing. New source + new Thing page + four updates.

  • gpt56-system-card (Report, source, T1, deploymentsafety.openai.com) — OpenAI's official system card (variants Sol/Terra/Luna): High in Cybersecurity + High in Biological/Chemical under the Preparedness Framework (Below High self-improvement; none Critical); cyber = finds vulns/exploit fragments but no autonomous end-to-end attacks (+ a noted agentic "exceeds user intent" regression); bio = 3/4 wet-lab-troubleshooting evals above High (novice uplift) but below Critical on protein design/novel pathogens; mitigations = activation classifiers that intervene mid-generation + 700k+ GPU-hr red-teaming; benchmarks HealthBench 60.5 (↑ from 51.8). States "government coordination was requested before broader deployment."
  • preparedness-framework (DefinedTerm) — OpenAI's internal capability risk-tiering as a frontier-lab self-governance instrument; contrasted with public standards (nist-ai-rmf, iso-iec-42001) — self-authored, self-graded, no independent auditor.
  • Updated openai-gpt56-access-restriction (primary-source corroboration section: capability + government-coordination now provider-documented; only the per-customer mechanism stays secondhand), export-controls-on-ai (open thread #1 — contested-facts problem easing on the capability side for GPT-5.6), eo-14409 (provider-side confirmation of the gate), and synthesis (new thread (c): the capability premise is provider-documented; lab's internal gate sits alongside the state's external gate; self-assessment caveat).
  • Routing: dominant substance = safety/preparedness/assurance (responsible-AI) → here; runner-up ../llm-providers-wiki (the model/benchmarks: Sol/Terra/Luna, HealthBench) noted as cross-spoke, not duplicated. Caveat recorded: T1 but a self-assessment (authoritative on what was tested, self-interested on framing). Entity openai reused cross-wiki; no new entity nodes.

[2026-07-03] ingest | Anthropic models revived — first completed EO-14409 cycle (Axios)

Msg 721 (axios.com/2026/07/03/anthropic-ai-models-revived-behind-the-scenes, via Telegram). WebFetch 403'd (Axios hard-block) → fetched via firecrawl. T2 (reputable outlet, named reporters Curi/Isenstadt/Gold; but load-bearing narrative is anonymous "U.S. official"/"source familiar" on a contested political story — weight the sequence over any actor's motive). This is the resolution chapter of the June export ban: Fable 5 / Mythos 5 back online 1 July 2026 after a 20-day showdown. New source anthropic-models-revived-2026 + 2 entities: caisi (GovernmentOrganization — the US frontier-model evaluator that, with the NSA, gated the release) and howard-lutnick (Person — Commerce Sec who made the 12 June call taking the models offline). Did not re-page Anthropic/Amodei/Fable-5 (canonical in ../research-wiki — cross-linked, not duplicated); Bessent/Brown/Heck/Cairncross named inline, not paged (entity cap + governance-relevance). Governance substance: first full ban→review→re-release loop of the eo-14409 regime, which revealed the regime is only half a process — defined entry (submit 30 days out), undefined discretionary exit (approval); the article's own bottom line is the approval "framework with transparency standards and timelines" still doesn't exist. Folded into synthesis (export-control instrument thread + a new open question on the approval half), eo-14409 (new "first completed cycle" section), anthropic-export-ban-2026 (resolution note — and the trigger got more contested via the experts' open letter). +3 pages. Ran avoid-ai-writing. Per-route verify deferred (routine ingest, no structural page move).

[2026-08-14] quality | research pass: the regime already in force

Coverage edge 4 hunted and closed. garante-openai-jurisdiction-ruling (Christakis & Monga, European Law Blog, T1, on Tribunale di Roma R.G. 4785/2025) supplies the regulation text (GDPR Art. 56 one-stop-shop) and an enforcement decision touching an AI system (the Garante's €15M ChatGPT fine, annulled on jurisdictional grounds 2026-03-18 — the merits untouched). Names a successor governance question: whether the one-stop-shop mechanism opens a jurisdictional gap for AI providers that incorporate in the EU after a DPA investigation starts. synthesis.md open questions + growth edge 4 updated; index.md gains one Article entry. Entity openai reused cross-wiki (canonical in ../llm-providers-wiki); no new entity nodes.

[2026-08-22] quality | Cycle: the first record of harm, and why it is not a rate

Coverage edge 6 half-closed. ai-incident-database (McGregor, IAAI-21, arXiv:2011.08512, T1) read in full: 1,000+ incident reports, numbered incidents joining multiple reports (incident 3 is the 737 MAX with 18), deliberately untitled records, namespaced competing taxonomies, and no top-down analysis. Built on the CVE identifier model rather than the NTSB investigation model — the paper says so directly.

The finding recorded in synthesis.md: this cannot serve as a denominator. Coverage follows press attention (the paper calls report count "a proxy for interest"), there is no severity scale, and submission is voluntary, so growth edge 1's certification-versus-outcomes question stays unanswerable. ../defensive-security-wiki hit the same wall with breach data — noted as a cross-spoke pattern, not merged.

Entity partnership-on-ai created, because the database's taxonomy design follows from its board's composition. index.md gains two entries; successor edge 6a written.

[2026-08-25] ingest | FDA AI/ML device adverse events — the denominator, missing under statute — via research pass

Model claude-opus-5, effort low. Edge 6a asked for a dataset carrying deployments as well as incidents, or a regulator's own statistics under a mandatory reporting duty. Medical AI is where that was most likely to exist: the FDA publishes the roster of devices it authorises, mandates reporting, and publishes the reports in MAUDE.

ai-ml-device-adverse-events (Babic, Cohen, Stern, Li & Ouellet, npj Digital Medicine 8:328, 2025, T1) works that data and reports the hole rather than filling it. 950 AI/ML devices authorised as of August 2024; 823 devices linked to 943 adverse-event reports, 2010–2023; and the authors' own sentence — "without an overall 'denominator' for the frequency of a medical device's use, it is not possible to talk about the relative safety of one product versus another."

Recorded as a half close, and the informative half. ai-incident-database's failure as a rate looked like the limits of a volunteer project. The same hole in a federal database built under statute makes it a property of the instrument. Also on the page: the reports are too thin to stratify — event location missing for every MDR in the sample, reporter profession missing in 73%.

Successor edge 6b written: every roster the corpus has found counts artefacts (devices, models, systems), and a rate needs episodes of use. Ingested url:-only with published: 2025-06-02; index.md gains one entry. Rung-0 checks only; the cycle verifies once at the end.

[2026-08-28] ingest | The AI Act's post-market monitoring duty (Arts 12, 19, 72) — research pass, T1

Regulation (EU) 2024/1689 of 13 June 2024, read from the Official Journal HTML (CELEX 32024R1689). Ingested with claude-opus-5 at high effort, hunting ## Most wanted #32 / this spoke's edge 6b — "a count of uses, not of things."

Created ai-act-post-market-monitoring (T1, the legislative text itself). Updated eu-ai-act, synthesis and index.md. Not a duplicate of eu-ai-act: that is a Thing page describing the law; this is a source page recording what the text says at three articles, which no page here had read.

Edge 6b is CLOSED, in the negative, and the negative is worth more than a half-close would have been. Article 72 requires data "on the performance of high-risk AI systems… which allow the provider to evaluate the continuous compliance" — performance and compliance, never volume. Deployer input is permissive ("may be provided by deployers"). Law-enforcement operational data is excluded from the obligation outright. And the monitoring plan is filed inside Annex IV technical documentation, which describes the system and is produced on request, not published.

The exception is the sharp part. Article 12(3) does mandate a per-use record — "recording of the period of each use of the system (start date and time and end date and time of each use)" — for remote biometric identification only (Annex III 1(a)). Article 19 then keeps those logs with the provider for at least six months, "to the extent such logs are under their control," with no reporting route. So the Regulation demonstrably can express a usage-counting duty, wrote one exactly once, for the most sensitive category, and left it private. That is a better answer to "is there a denominator anywhere" than another roster of artefacts would have been.

Recorded with a hedge, deliberately. Article 72(3) required the Commission to adopt an implementing act with the monitoring-plan template by 2 February 2026; searching on 2026-08-28 found no adopted act and several compliance write-ups calling it pending. That is not an official negative and the page says so. What is certain from the Regulation: the high-risk obligations applied from 2 August 2026.

Successor edge 6c written: billing produces episode counts as a by-product, so a health regulator may answer where this one does not — needs an FDA post-market or NHS/CMS utilisation figure for an authorised AI device. Cooldown 2026-09-11.

Entity note: no new nodes. The European Commission and the Union legislature are context for a text already paged as eu-ai-act; nothing here rests on a claim about either as an actor.

[2026-08-30] ingest | CRS on EO 14409 — the US instrument is voluntary (quality cycle, queued re-read)

us-ai-policy was the page the freshness queue kept flagging and the cycle kept deferring — a living policy survey, last updated 28 June. Read at last, and it produced this spoke's largest correction.

Ingested crs-eo-14409-explained (CRS In Focus IF13268 v2, 9 July 2026, T1; PDF read with pypdf after congress.gov 403'd a plain fetch) and eo-14365 (the December 2025 preemption order, T1 from the White House text — previously referred to on us-ai-policy with no number).

The correction. eo-14409 does not require anything. CRS: companies "are asked to participate in a voluntary review process"; the order's provisions "stop short of formal licensing or requiring preclearance." It does not define covered frontier model — that was a deliverable due 1 August 2026 from a classified benchmarking process. Nine tasks, all on agencies. And CRS lists the Anthropic suspension among the administration's other approaches, in contrast to the EO, not as enforcement of it.

Both readings kept per record-don't-overwrite. Ours is worse-sourced and the sequence it rests on is still unexplained: two frontier labs restricted within a fortnight of an order about frontier-model security.

Pages: new crs-eo-14409-explained, eo-14365. Updated eo-14409 (the requirement claim retracted in place with the superseded text kept; the "enforcement record" section reframed), us-ai-policy (both poles corrected — EO 14365 named with its three levers, EO 14409 downgraded from gate to request, plus a new section on what is not there: no federal AI statute, two bills unpassed), synthesis.md, index.md.

The methodological point, folded into synthesis. Nothing here was fabricated; the wiki read a sequence as a mechanism. A voluntary framework and a mandatory gate look identical from outside when everyone participates and two participants are restricted at once. Rule taken: when a page says an instrument compels something, the obligation is the claim to check against primary text.

Also new to the corpus: NSPM-11 (5 June 2026), pointing the opposite way — make frontier models broadly available to national security professionals; and NIST already had pre-release access to Anthropic's and OpenAI's models by private agreement, before any order, with no public transparency.

[2026-09-06] ingest | watermarks-remover — the other end of a labeling mandate

Routed here by the hub from Telegram. An MIT-licensed Python project (GitHub, read via the API the same day: created 2026-08-11, 20,845 stars / 2,404 forks, v0.7.0) that strips AI provenance marks — invisible Unicode carriers, hard-bound C2PA/EXIF/XMP across ~25 container formats, and best-effort rewrites against statistical text watermarks. Ships as a Claude Code plugin, an agent skill and a local HTTP service.

Graded T1, which needs saying because the ladder's first instinct for a project README is T3. The load-bearing claims here are artifact fact — coverage matrix, format table, licence, repo metadata — plus three statements that cut against the project: no tool can certify a vendor detector will fail; removing a text watermark means rewriting the text and capping it at the rewriting model's ceiling; and C2PA soft bindings survive a metadata strip. Nothing in it is an efficacy claim, so there is no maker-speaking claim to demote.

Why it belongs in a governance wiki. china-ai-regulation already recorded mandatory embedded provenance metadata + watermarking as the hard-law end of the provenance spectrum. This is what the other end of that duty costs. The duty binds the producer at generation time and reaches nobody downstream, so the instrument's bite depends on detection — which is vendor-held, in private preview (Anthropic's text detector) or, per this source's coverage matrix, retired (Google, August 2026). That last one is single-sourced and flagged as such on the page.

Pages: new watermarks-remover, content-credentials-c2pa (the file-level provenance mechanism; no page anywhere in the corpus covered C2PA), guillaume-meyer. Updated china-ai-regulation, synthesis.md (new open question on whether a marking duty survives free removal tooling; new growth edge 6d on whether removal is regulated anywhere; two new cross-spoke adjacencies), index.md.

Not duplicated: ../llm-inference-wiki owns the token-level mechanism (text-watermarking, claude-text-watermark) and this ingest links to it rather than restating it. ../agentic-tooling-wiki owns the packaging half (skills, plugin marketplace, the skills-vs-hooks determinism argument), noted as cross-spoke context on the source page.

[2026-09-16] ingest | "Anthropic's mother of all risk factors" (FT)

Source: anthropic-mother-of-all-risk-factors, T2 (source) — unread beyond the headline: ft.com refused WebFetch, curl hit a bot check, archive mirror blocked. Substance held from The State of AI (2026-09-09) and Palihapitiya coverage (2026-09-11), marked as such on the page. Routed by the hub; runner-up llm-providers (Anthropic as provider / IPO). No new entities (Anthropic lives in research-wiki). Synthesis section The risk statement acquires liability.

[2026-09-27] research | edge 6c — a count of uses

Source: ai-device-cpt-claims, T1 (source; read at abstract level, NEJM AI 403 and medRxiv bot check). Closes edge 6c: billing codes give a usage count for AI devices, and only two classes pass 10,000 claims. The regulator still collects none. Residual written as edge 6e (divide harm reports by use). No new entities. Synthesis edge 6c updated in place.

[2026-09-27] quality | quality cycle

Scorecard in the hub quality-log.md. Integrity clean. One research-pass page added (above).

[2026-10-07] research | edge 23/6b — checked Article 73 (serious-incident reporting)

Source: ai-act-post-market-monitoring refreshed in place (same page, now also Article 73), T1, read directly from the Official Journal text at eur-lex.europa.eu (two WebFetch passes to reach the right offset in a 525k-character document). Article 73 is the Regulation's one duty with real deadlines (15/2/10 days by severity) and it is incident-by-incident, no aggregate field. Confirms the closed edge 6b finding extends to the Regulation's other enforced duty, not just Article 72. Also recorded a named, independent statement of the same gap outside the corpus: Santoro (Tech Policy Press, 2026-04-24) calls it "the denominator problem" and advises an AIID/MIT FutureTech/Arcadia Impact project estimating deployment scale because no regime supplies it directly.

[2026-10-11] research | a method for the missing denominator (edge 23, half-close) | ai-incident-public-health-lens

Research pass hunt. McGregor (AIID author) + 7 co-authors propose a public-health-surveillance method for converting AI incident counts into rates, using an exposure denominator — worked case: autonomous vehicles, California DMV miles-driven. Updated ai-incident-database. Half-close: the method works where mandatory reporting exists; general-purpose AI models still have no published exposure denominator.

Related pages

Links to