The sources
Where each statement on this site comes from
The claims register of the presenter pack, as published. Confidence: firm, a primary source checked; sourced, a reliable secondary source; ours, the Hub’s own reading or proposal; check, cited in working papers and to be verified before quoting.
Research and market
| # | Claim | Where it appears | Source | Scope and date | Confidence | Say it like this | Do not say |
|---|---|---|---|---|---|---|---|
| 1 | The harness explained 7.8 times more of the variance in results than the model | Part three, essential 1 | Zhang, Wang, Ge, Xu, Hamm, Reddy, Stop Comparing LLM Agents Without Disclosing the Harness, arXiv 2605.23950, 7 May 2026, section 4.2 | harness-induced against model-induced variance, 18.48 against 2.37 pp², on a 100-task subset of SWE-bench Verified with three near-tied frontier models and three harness configurations; the authors say the ratio is not universal | firm | "In a controlled study on coding tasks, the harness mattered several times more than the model. So a firm approves a release, not a model." | that it was measured on financial decisions, or that the model does not matter |
| 2 | No product we found keys a decision record to both the mandate version and the build | Part three, essential 5 and the three joins | the Hub's vendor review, public product documentation, September 2026 | 23 vendors; documentation, not tests | ours | "In the products we reviewed in September, we found none that does this; Credo AI's Agent Governor records the policy version in force, in research preview." | that no product can do it, or that we tested the products |
| 3 | Vendor placements on the map (product sold for this; feature inside a broader product; named, not seen shipped) | Explorer, "Who sells what" and the vendor lens | the Hub's vendor review, September 2026 | public descriptions; an initial reading | ours | "These are placements from public documentation, to check against the version a firm runs. They are not an assessment." | that a vendor is compliant, certified, or ranked |
| 4 | Narrow or widen, the step that moves the mandate on evidence, appears in none of the products reviewed | Explorer, "Who sells what" | as row 3 | as row 3 | ours | "None of the products we reviewed moves the mandate on evidence; that return path is one of the three joins a firm builds." | that it cannot be bought in future |
| 4a | No product we found reads the standing of the basis at the moment of commit, which is what joins the second essential to the fourth essential | Chapter 05, what you govern when you govern claims | the Hub's vendor review, public product documentation, September 2026; the knowledge-plane products read against the check's six inputs | 23 vendors and the knowledge-plane category; documentation, not tests | ours | "In the products we reviewed in September, management planes inventory the agents, control planes gate the tool call, and knowledge planes hold what the agent knows; we found none that reads the standing of the basis when the act commits." | that no product can, or that we tested them |
| 4b | None we found sells either rule: autonomy may not rise above evidence, or standing scales with consequence | Chapter 05, the second rule; part three, essential 6 | as row 4a | as row 4a | ours | "We found no product that decides either rule for a firm; both are the firm's to write, and the products supply inputs to them." | that the rules are a product feature, or that a vendor is missing them |
| 5 | Each platform brings its own control plane, for its own agents | Part one, the question; part four, the estate | vendor documentation (Microsoft Agent 365 and Entra Agent ID; Google Agent Gateway, Agent Identity, Agent Registry; AWS Bedrock AgentCore) | September 2026 | sourced | "Each platform governs its own agents. Nobody owns the identity across them unless the firm does." | that any platform is insufficient for its own agents |
| 6 | Rules-based sanctions screening produces mostly false positives | Payments case | vendor statements | vendors cite around 99% | check | "Most hits are false positives; vendors cite figures near 99%. The case's own rates are assumptions." | a precise industry rate |
| 6a | Documented AI incidents rose from 233 in 2024 to 362 in 2025; the adoption of AI reached 88 percent of organisations surveyed in 2025 | The opening, capability and control | Stanford HAI, AI Index 2026, 14 April 2026: the responsible AI chapter (incidents, from the AI Incident Database) and the economy chapter (adoption, from the McKinsey Global Survey on AI 2025: 1,993 respondents in 105 countries, fielded June to July 2025, published 5 November 2025) | incidents as reported to one public database, so a count of what was documented, not of what happened; adoption as regular use in at least one business function | firm | "Documented incidents went from 233 to 362 in a year, up by more than half, while adoption reached 88 percent. The database counts what was reported." | that incidents rose 55 percent in the world, or that 88 percent of firms run agents |
| 6b | One firm in five among those planning agents within two years, 21 percent, has a mature governance model for them | The opening, capability and control | Deloitte, The State of AI in the Enterprise: The Untapped Edge, 21 January 2026; 3,235 business and IT leaders in 24 countries, director to C-suite, fielded August to September 2025; the 21 percent is of the respondents planning to deploy agentic AI within two years | self-reported maturity, on agentic AI specifically, among the firms planning it | firm | "Among the firms planning agents within two years, 21 percent said they had a mature governance model for them, in a survey of 3,235 leaders in 24 countries. In Deloitte's words, agents are scaling faster than their guardrails." | that 21 percent of all firms have one, that four in five firms have no governance, or that the figure was measured |
| 6c | The next agents carry a world model and plan an act by predicting its consequences ("like JEPA"); the map does not move for that | Part one, what the map does not draw yet; the research page | LeCun, A Path Towards Autonomous Machine Intelligence, OpenReview, 27 June 2022 (the position); Assran et al., I-JEPA, CVPR 2023, pp. 15619–15629; Bardes et al., V-JEPA, TMLR 2024; Assran et al., V-JEPA 2, arXiv 2506.09985, June 2025 (the line of work) | a research direction, stated by its author and shown on images, video, and robot planning; no agent in a regulated decision runs on it today | sourced | "The next generation of agents is being built to predict the consequences of its own acts, the world-model line of work, like JEPA. For the map that changes nothing: a predicted consequence is one more claim the check reads, with a standing." | that LLM agents plan this way today, that JEPA is a product, or that a world model makes the check unnecessary |
| 6d | NIST's AI Risk Management Framework organises AI risk work in four functions: govern, map, measure, and manage | The reference model, the frameworks | NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023 | a voluntary framework; US, used internationally | firm | "NIST's framework is a risk workflow: govern, map, measure, manage." | that it is binding, or that following it meets the AI Act |
| 6e | ISO/IEC 42001 is a management system standard for AI: policy, roles, evidence, review, and improvement, certified and audited like other management systems | The reference model, the frameworks | ISO/IEC 42001:2023, Information technology - Artificial intelligence - Management system, December 2023 | an international standard; certification is voluntary | firm | "ISO/IEC 42001 makes AI management a management system, audited like the others." | that certification to it shows an agent is safe to act |
| 6f | In networking, the control plane instructs devices how to process and forward traffic; the management plane monitors, configures, and maintains them | The reference model, the plane strictly; part one, the question | IETF RFC 7426, Software-Defined Networking (SDN): Layers and Architecture Terminology, January 2015, the terminology section | a definition from networking, borrowed by analogy | firm (read 12.10.2026) | "Networking defined the term first: the control plane instructs, the management plane monitors and configures." | that a standard defines the AI control plane |
| 6g | The autonomy levels A.0 to A.5 follow Kasirzadeh and Gabriel's agentic profiles | The reference model, the rule; part three, essential 6; the research page | Kasirzadeh A., Gabriel I., Characterizing AI Agents for Alignment and Governance, arXiv 2504.21848, 30 April 2025; published in Nature in August 2026 according to the reference model outline | the levels are theirs; the evidence levels and the rule are the Hub's | sourced (the preprint read 12.10.2026; check the Nature version before citing it) | "The autonomy levels follow Kasirzadeh and Gabriel; the evidence levels and the rule between them are our proposal." | that the authors endorse the rule or the evidence levels |
| 6h | Kubernetes admission control intercepts a request after it is authenticated and authorised and before the object is stored, and can reject it on policy | The reference model, the plane strictly | Kubernetes documentation, Admission Control in Kubernetes (kubernetes.io) | the API server's admission phase, as documented | firm | "Kubernetes admission control can still refuse a request that is authenticated and authorised, on policy, before it is written." | that every architect uses it, or that it governs AI agents |
The cases and their figures
| # | Claim | Where | Source | Confidence | Say it like this | Do not say |
|---|---|---|---|---|---|---|
| 7 | The use cases, their firms and the city, the claims, payments, conversations, and applications, and their figures | Part two, the dossier, the explorer | invented for illustration | ours | "These cases are invented. What a real workshop marks is the firm's own workflow." | |
| 7a | The city case's guidance versions, their content, and the allowance amounts | Part two, the city case | invented for illustration | ours | "The guidance and the amounts are invented. The point is that the law as read changes on a known date, and the check asks which version applies." | |
| 8 | Business case: 96,000 claims, 25 minutes saved, €60 an hour, 65% realised, €0.6m to run, €0.9m to set up, about €0.96m a year | Part two, the decision; part four, the value | the motor case pack | ours | "Illustrative figures; the point is the chain, and what happens to it when less of the saving is realised." | |
| 9 | Thresholds: three referrals in seven days, restored after 20 clean proposals, a weekly evaluation sample | Part two, the loop; the explorer | the case packs | ours | "The numbers are illustrations of a rule written in advance. A firm sets its own." | |
| 9a | Curation costs half to one person per domain once it runs, a domain being one product line's wordings, one list service, or one legal register; the first load is the pilot's own cost | Chapter 05, how knowledge is prepared | the Hub's own estimate from the Governed Intelligence work, not measured in a deployment | ours | "An illustration, from our own reading: budget in that order per domain, after a first load the pilot measures. It is not a benchmark, and the run rate is not the backlog." | that a bank with five hundred product lines needs five hundred people; the domain is the unit, and a firm counts its own |
| 9b | Most programmes we have seen stop at a library and a search index, buy capture and structure and call it done, have not staffed curation, and have not written the e2 charter | Chapter 05: the opening, the jobs that share one word, and how knowledge is prepared | the Hub's reading of its own client work and the programmes it has reviewed, 2025 to 2026; not surveyed, not counted | ours | "In the programmes we have seen, the knowledge work stops at retrieval; we have not counted it." | "most firms" as a market statistic; any percentage |
| 9c | Most pilots we have seen keep K0 evidence, logs without attribution; most firms we have seen can show K1, which licenses A.1 and A.2 | The opening, the ladder; part three, essential 6; the reference model, the levels and the rule | the Hub's reading of its own client work and the programmes it has reviewed, 2025 to 2026; not surveyed, not counted | ours | "In the programmes we have seen, pilots log without attribution, and the firms that have gone further can show every act recorded and attributed." | a share of the market, or "most firms" as a statistic |
Law and supervision
| # | Claim | Where | Source | Confidence | Say it like this | Do not say |
|---|---|---|---|---|---|---|
| 10 | SR 26-2 replaced SR 11-7 and puts generative and agentic AI out of its scope | Reference model, supervisory annex; the reference model page, slide 1 | SR 26-2, Federal Reserve, OCC, and FDIC, 17.04.2026; summaries by CIMCON, Management Solutions, Moody's | sourced: quote the footnote from the primary text | "The US model risk guidance of April 2026 leaves agentic AI out of its scope and tells banks to govern it with their own risk management." | that it exempts agents from control, or that it applies in the EU |
| 11 | Motor claims are not high-risk under the AI Act; a collections conversation is arguable | Part four, the law; the motor case | Regulation (EU) 2024/1689, Annex III point 5: (b) creditworthiness of natural persons, (c) risk assessment and pricing for natural persons in life and health insurance | firm | "Annex III lists credit scoring of individuals and life and health pricing; motor claims are not on it. A collections agent that weighs affordability is arguable, so the bank records its reasoning." | that the AI Act does not matter to insurers |
| 11a | The AI Act's duties depend on the use and on the firm's role, provider or deployer | The reference model, the frameworks | Regulation (EU) 2024/1689, Article 3(3) and (4) (the definitions), Article 16 (providers of high-risk systems), Article 26 (deployers), Article 6 and Annex III (the high-risk uses) | firm | "The Act sets duties by the use and by the firm's role: a provider's duties differ from a deployer's." | that a firm is only ever one of the two, or that every agent is high-risk |
| 12 | Article 22 GDPR applies to a decision based solely on automated processing with a legal or similarly significant effect | Part four, the law; the dossier | Regulation (EU) 2016/679, Article 22; CJEU C-634/21 SCHUFA (07.12.2023) | firm | "Whether a settlement falls under Article 22 is a test the firm applies, decision type by decision type." | that every agent decision is covered, or none is |
| 13 | For an automated decision within Article 22, the explanation must describe the procedure and principles actually applied | Part four, the law; part three, the record | CJEU C-203/22 Dun & Bradstreet Austria, 27.02.2025, on Article 15(1)(h) | firm | "Where a decision falls under Article 22, the court asks for the procedure and principles actually applied, in a form the person can understand. A decision record keyed to its versions can produce that." | that every decision must be explained this way, or that the firm must disclose its model |
| 14 | For instant euro payments, EU-list sanctions screening moves to the bank's own customers, checked daily and after every new measure | Payments case notes and lens | Regulation (EU) 2024/886, Article 5d | sourced | "For instant euro payments, banks check their customers daily instead of screening each payment; our case is a payment to a bank outside the EU, screened as it leaves." | that the rule covers payments to banks outside the EU |
| 15 | EBA guidelines ask for trained analysts and two people for higher exposure | Payments case | EBA/GL/2024/15, paragraphs 32(d) and 33 | sourced | "The guidelines want alerts analysed by trained staff, and two people above an exposure. Automated clearance in approved categories is the bank's own reading, written into its screening policy." | that the guidelines permit automated clearance |
| 16 | The Consumer Credit Directive II gives consumers a right to human intervention, an explanation, and a review, and asks for reasonable forbearance before enforcement, for agreements concluded from 20 November 2026; earlier agreements stay under Directive 2008/48, which since 30 December 2023 carries a forbearance duty of its own (Article 16a, inserted by Directive 2021/2167) | Part four, the law; credit and collections cases | Directive (EU) 2023/2225, Articles 18, 35, and 47; Directive 2008/48/EC, Article 16a | sourced (Article 16a checked by a second reader, 10.10.2026) | "For credit agreements concluded from 20 November 2026, a consumer can ask for a person, an explanation, and a review, and lenders show forbearance before enforcement. Older agreements already carry a forbearance duty under the 2008 directive." | that it applies to lending to companies, or that older loans carry no forbearance duty |
| 17 | EIOPA's opinion on AI governance and risk management sets expectations for insurers using AI outside the AI Act's high-risk list; it is addressed to national supervisors and binds no one on its own | Part four, the law | EIOPA, Opinion on artificial intelligence governance and risk management, 6 August 2025 | sourced (date and addressees checked by a second reader, 10.10.2026) | "EIOPA's opinion of August 2025 expects proportionate governance, data quality, oversight, and explanations from insurers using AI, in a risk-based way." | that it is binding, or that it covers high-risk systems, which the AI Act does |
| 18 | AI literacy: providers and deployers take measures to support the AI literacy of their staff, without having to guarantee any level for any individual | Part four, the law | AI Act, Article 4(1) as reworded by Regulation (EU) 2026/1744 (the omnibus) | sourced (the reworded text read on EUR-Lex by a second reader, 10.10.2026; quote it before publishing) | "The AI Act asks firms to support the AI literacy of the people who use it; since the 2026 omnibus the duty stays, and no particular level has to be guaranteed." | that the Act requires a training programme, or that the omnibus removed the duty |
| 19 | A German works council co-determines technical systems able to monitor employee performance | Part four, the people | Betriebsverfassungsgesetz §87(1) no. 6 | firm | "In Germany the works council has a say before a system that can monitor performance is introduced; the decision record names handlers." | that the works council can veto the agent itself |
| 20 | The FSB declined continuous human monitoring (10.06.2026) and Breeden said the same (30.06.2026) | Reference model outline, A.4 | cited in the reference model outline v02 | check | do not quote until checked against the primary texts | |
| 20a | AI systems intended to be used by or on behalf of public authorities to evaluate the eligibility of natural persons for essential public assistance benefits, or to grant, reduce, revoke, or reclaim them, are high-risk unless Article 6(3) takes them out; after the 2026 omnibus these obligations apply from 2 December 2027 to systems put into service from then; a system already in service follows on its next significant change in design, and one used by a public authority by 2 August 2030 at the latest | Part two, the city case; part four, the law | Regulation (EU) 2024/1689, Annex III point 5(a), Article 111(2); Regulation (EU) 2026/1744 of 8 July 2026 (the AI omnibus), published in the Official Journal on 24.07.2026, in force 27.07.2026 | firm for the list; sourced for the dates (quote the Official Journal before publishing) | "An AI system a public authority uses to decide an essential benefit is high-risk under the AI Act. After this year's omnibus, those obligations apply from 2 December 2027 to new systems; one already running follows on its next significant change, and by August 2030 at the latest." | that the obligations apply today, that the omnibus removed them, or that a system already in service escapes them |
| 20b | A public body deploying such a system has duties that include human oversight by trained people, logs kept for at least six months, telling the people it decides about, a fundamental-rights impact assessment before first use, and registering its use in the EU database | The city case notes and its regulation lens | AI Act Articles 26, 27, and 49 | check: confirm against the text as amended by the omnibus | "A city that deploys it assigns trained people to oversee it, keeps the logs, tells residents, assesses the impact on their rights before it starts, and registers the use." | article numbers or periods after the omnibus without checking |
| 20c | In Germany, a benefit decision may be issued fully automatically where there is no reason for an official to handle the case, taking into account facts from the applicant the procedure cannot evaluate; in general administrative law only where a statute permits it and there is no discretion or margin of appreciation | The city case, its regulation lens and notes | § 31a SGB X; § 35a VwVfG | firm | "Whether a machine may decide at all is national law. Germany allows it for benefits where no official needs to look at the case, and elsewhere only where a statute allows it and leaves no discretion." | that the rule is EU-wide, or that German law forbids automated benefit decisions |
| 20d | Australia's Robodebt scheme raised welfare debts by averaging annual income over fortnights, was found unlawful, and was the subject of a Royal Commission that reported in July 2023 | The city case notes | Royal Commission into the Robodebt Scheme, report, July 2023 | sourced | "Automated reclaims at scale are where benefit systems have done the most harm; Robodebt is the case everyone in the public sector knows." | figures for the debts or the settlement without checking |
| 20e | A German benefit notice states the essential factual and legal reasons; where the Land keeps the objection procedure, an objection within one month is decided by the objection authority, and the administrative court is open after it or directly where the Land has abolished the objection; housing allowance disputes go to the administrative courts | Part four, the law, the city case | § 35 SGB X; §§ 68 to 70 VwGO; § 40 VwGO and the absence of housing allowance from § 51 SGG; Land laws abolishing the objection in some matters | sourced (second reader, 10.10.2026; a public-law reader should confirm before external use) | "The notice gives its reasons, and the household can object within a month, or go to the administrative court; a person decides the objection, never the machine." | that the objection is decided by the same caseworker, or that the procedure exists in every Land |
| 20f | Under Article 22(2)(b) GDPR a decision authorised by Member State law escapes the ban on solely automated decisions only where that law itself lays down suitable safeguards; the safeguards of Article 22(3) attach to (2)(a) and (c) | Part four, the law, the city case | Regulation (EU) 2016/679, Article 22(2)(b) and (3), recital 71 | firm | "For a benefit decision authorised by national law, the safeguards come from that law: here the duty to weigh what the applicant states, the reasons in the notice, and the objection." | that Article 22(3)'s trio applies by itself to a decision under (2)(b) |
| 20g | Article 50 of the AI Act applies since 2 August 2026 to an AI system that interacts directly with natural persons; the 2026 omnibus did not postpone it | Part four, the law, every case where the agent answers the customer | Regulation (EU) 2024/1689, Article 50(1) and Article 113; Regulation (EU) 2026/1744 | sourced (second reader, 10.10.2026) | "Where the agent answers the customer itself, it says it is an AI system; that duty is in force since August 2026." | that the omnibus delayed Article 50, or that a corporate client's staff are not natural persons |
| 20h | The AML Regulation applies from 10 July 2027; until then the AML duty is Directive 2015/849 as transposed; the Regulation brings the check of customers against targeted financial sanctions into customer due diligence | Part four, the law, the payment case | Regulation (EU) 2024/1624, Articles 3 and 20(1)(d), and its date of application; Directive (EU) 2015/849 | sourced (second reader, 10.10.2026) | "Today the bank's AML duty is the directive as transposed; from July 2027 the regulation applies, and checking customers against the sanctions lists becomes part of due diligence." | that the AML Regulation applies today |
| 20i | The EBA guidelines on restrictive measures address institutions the EBA supervises and payment and crypto-asset service providers, not insurers; the sanctions regulations themselves bind every firm that holds or moves funds | Part four, the law, the motor case | EBA/GL/2024/14 and EBA/GL/2024/15, applying from 30 December 2025; Council regulations on restrictive measures | sourced (second reader, 10.10.2026) | "The EBA guidelines are written for banks and payment firms. An insurer is bound by the sanctions regulations themselves, so claim payees are screened." | that an insurer is outside sanctions law |
| 20j | DORA reaches identity management and access control, p3, through its technical standard on the ICT risk management framework | Part four, the law; the explorer's DORA lens | Regulation (EU) 2022/2554; Commission Delegated Regulation (EU) 2024/1774, the RTS on the ICT risk management framework, its articles on identity management and access control | check (raised by the third panel's CISO, 11.10.2026; verify the article numbers before quoting) | "DORA's risk management standard covers who may access what, so agent identity and access sit under it as well as incidents, testing, and third parties." | the article numbers until read |
Statements about the set itself
| # | Claim | Say it like this |
|---|---|---|
| 21 | What the map is | "A proposal for discussion from the Hub's research. It is not a product, and it certifies nothing." |
| 22 | What the control plane is | "It instructs, and can say no before the act. Platforms already instruct: identity, routing, guardrails, access. The part that refuses is the mandate and the check, joined to the record and the loop. The map is everything the plane needs to be right." |
| 23 | The reference model and its levels | "Autonomy levels A.0 to A.5 and evidence levels K0 to K3 are a proposal under review, with a first version planned for the fourth quarter of 2026." |
| 24 | Who wrote it | "Hub research by Witold Reichhart." |