Proof Before Autonomy

Can automated compliance be trusted? The published evidence says: not on its word. In 2024 general-purpose language models hallucinated on 58% to 88% of verifiable questions about US federal court cases, and the purpose-built legal research tools sold by LexisNexis and Thomson Reuters still hallucinated on 17% to 33% of a preregistered set of queries. By late 2025 three legal AI products scored 78% to 81% on a research benchmark where lawyers scored 71%, so on that benchmark the machines came out ahead, but none of these numbers is a threshold anyone has agreed to. Regulators in London, Paris, Washington and Brussels have, meanwhile, said the same thing in different words: a named person stays accountable, the system must be overseen, its logs kept, its data located. Industries that already run safety-critical software on autopilot got there a different way. Aviation certifies software to a level set by the harm its failure could cause and reports 1.32 accidents per million flights; the US Food and Drug Administration clears AI devices one intended use at a time and has authorised more than 1,600 of them. This article assembles that evidence and ends with four conditions, each traceable to a rule or a measurement, that any automated compliance system should meet before it is allowed to act alone.

In this piece: how accurate the machines are, really; what regulators have said about who answers; how aviation, medicine and driving earn autonomy; evidence that runs both ways; where the data sits; the four conditions.

How accurate the machines are, really

The case for automating compliance rests on a claim about reading: that a machine can read a rule, a policy or a transaction record and say what it means with the reliability of a trained professional. That claim can be tested, and in 2024 two teams of researchers from Stanford’s RegLab and Yale did so on legal text.

The first study ran more than 800,000 queries about real US federal court cases through four general-purpose chatbots. On questions with a checkable answer the models hallucinated, meaning their output was “not consistent with legal facts”, “between 58% of the time with ChatGPT 4 and 88% with Llama 2”.¹ A model that is wrong more often than right on questions with a known answer is not a junior associate; it is a confident stranger.

The second study tested the products built to fix this: Lexis+ AI, Westlaw’s AI-Assisted Research and Thomson Reuters’ Ask Practical Law AI, all of which retrieve real documents before the model writes. Against 202 preregistered queries the three tools “each hallucinate between 17% and 33% of the time”, and the authors concluded that “the providers’ claims are overstated”.² The best tool answered 65% of queries accurately; the weakest, 20%, with 63% of its answers incomplete (refusals or ungrounded responses). Retrieval reduced the fabrication; it did not remove it.

Benchmarks that break legal work into kinds of reasoning show why a single accuracy number misleads. LegalBench, a 162-task benchmark whose tasks were “designed and hand-crafted by legal professionals”, scores GPT-4 at 89.9 on rule-conclusion (apply a stated rule to given facts) and at 59.2 on rule-recall (state what the rule is).³ The same model scores 30.7 points lower on one kind of legal reasoning than on another. Any system that automates compliance performs distinct tasks: classify a clause, map it to a control, draft a policy, flag a trade. The evidence says each must be measured on its own, because a score on one tells little about the next.

More recent results are more favourable to the machines. In a report published in October 2025, Vals AI, a benchmarking firm, put 200 US legal research questions to three legal AI products, to ChatGPT and to lawyers supplied by a US law firm. On accuracy the lawyers scored 71%; Alexi, Counsel Stack and Midpage scored 80%, 81% and 78%.⁴ No product from Thomson Reuters or LexisNexis, whose tools the second Stanford study tested, was among those evaluated, so the two results cannot be joined into one line. Harvey, a legal AI vendor, reports its own models producing “74% of a final, expert lawyer-quality work product” on its own BigLaw Bench, with hallucinations scored negatively.⁵ These are vendor-adjacent numbers and should be read as such. Taken together with the Stanford work, they describe a technology whose measured accuracy on narrow legal research tasks ranges from 20% to 81% across the products and studies cited here, on different definitions, and that still has no agreed pass mark.

One further finding bears directly on compliance, where the input can be wrong before the machine sees it. Asked questions built on a false legal premise, GPT-4 went along with the premise 69% of the time, and PaLM 2 99%.¹ A compliance system that accepts the firm’s own description of itself without challenge will faithfully automate the firm’s mistakes.

What regulators have said about who answers

Regulators have not waited for the benchmarks. Their statements since 2023 share one structure: the technology may be new, the accountability is not.

In the United Kingdom the FCA’s AI Update of April 2024 points to the Senior Managers and Certification Regime, which it says “emphasises senior management accountability and is relevant to the safe and responsible use of AI”. In dual-regulated and Enhanced SM&CR firms, one or more Senior Managers must “have overall responsibility for each of the activities, business areas, and management functions of the firm”, so “any use of AI in relation to an activity, business area, or management function” falls within a Senior Manager’s responsibilities, and every Senior Manager must “take reasonable steps to ensure that the business of the firm, for which they are responsible, is effectively controlled”.⁶ The regulators consulted on whether AI needed its own named role; of the 54 responses to the discussion paper DP5/22, most said a new Prescribed Responsibility for AI would not help.⁶ The accountable person, in other words, is whoever already owns the function the machine now performs.

The European Securities and Markets Authority said the same in May 2024, in the sentence that opens this article: “firms’ decisions remain the responsibility of management bodies, irrespective of whether those decisions are taken by people or AI-based tools”. It added that the management body “should have an appropriate understanding of how AI technologies are applied and used” and that ESMA “expects investment firms to maintain comprehensive records on AI utilisation”.⁷

In the United States FINRA’s Regulatory Notice 24-09 of June 2024 reminded broker-dealers that it “intends for its rules and guidance to be technologically neutral”, that obligations apply to generative AI “just as they apply when member firms use any other technology or tool”, and that a firm “must have a reasonably designed supervisory system tailored to its business”.⁷ The SEC’s examination priorities for fiscal 2025 state that the Division “will review registrant representations regarding their AI capabilities” for accuracy and “will assess whether firms have implemented adequate policies and procedures to monitor and/or supervise their use of AI”.⁷

The European Union wrote the oversight requirement into statute. Article 14 of the AI Act requires high-risk systems to be built so “that they can be effectively overseen by natural persons” who are enabled “to properly understand the relevant capacities and limitations of the high-risk AI system”, “to remain aware of the possible tendency of automatically relying or over-relying on the output”, and “to decide, in any particular situation, not to use the high-risk AI system” or “to otherwise disregard, override or reverse the output”. Article 26 obliges the firm that deploys such a system to “assign human oversight to natural persons who have the necessary competence, training and authority”, to monitor it, and to keep its automatically generated logs for “at least six months”.⁷ Whether a compliance tool counts as high-risk under the Act depends on its use; the design principles are the point here.

US bank supervisors have set out how a model should be validated, and in April 2026 they rewrote that guidance. The Federal Reserve, the FDIC and the OCC replaced SR 11-7, the model-risk guidance issued in 2011, with revised guidance (SR 26-2) that defines a model as “a complex quantitative method, system, or approach that applies statistical, economic, or financial theories to process input data into quantitative estimates”. It describes “effective challenge” by “objective experts”, validation of “conceptual soundness”, “ongoing monitoring” and “outcomes analysis”, which “compares model outputs to corresponding real-world outcomes”, and notes that “inherent risk increases with model complexity”. It also states: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.”⁷ A language model reading a rulebook is therefore outside the US agencies’ model guidance, which leaves such tools to a bank’s own “risk management and governance practices”; the tests the guidance describes, drawn from 15 years of supervisory experience, show what proving one could look like.

How aviation, medicine and driving earn autonomy

Finance is not the first industry to put software in charge of decisions that can hurt people. Three others did it earlier, and each built a mechanism by which the software earns its task rather than being granted it.

Aviation grades software by the harm its failure could cause. The FAA’s Advisory Circular 20-115D accepts the industry standard DO-178C as “an acceptable means of compliance for the software aspects of type certification”, ties the rigour demanded of each piece of software to a safety analysis that “assigns the minimum development assurance level based on the severity classifications of failure conditions”, and expects applicants to “satisfy all of the objectives associated with the software level assigned to the software”.¹¹ The industry’s outcomes are also on public record: IATA counts 1.32 accidents per million flights in 2025, 51 accidents across 38.7 million flights, with an all-accident rate that has stayed between 1.06 and 1.61 for a decade.⁸ The figure is not zero, and 2025 saw 394 on-board deaths; the industry publishes the number every year rather than declaring the problem solved.

Medicine clears software one intended use at a time. The FDA says it has authorised “over 1,600 AI-enabled medical devices” for US marketing as of September 2026, each through a premarket pathway, and that the devices on its list had a review including “an evaluation of study appropriateness for the device’s intended use and technological characteristics”.¹⁰ A peer-reviewed count of the list to the end of 2025 shows the pace: 42 devices authorised in the 21 years to 2015, 190 in 2016 to 2019, 406 in 2020 to 2022 and 792 in 2023 to 2025, 331 of them in 2025 alone.⁹ A device cleared to flag a suspected stroke on a CT scan is not thereby cleared to read a mammogram. The task, not the technology, is what gets authorised.

Driving publishes the failures. Since 2014 California has required manufacturers testing autonomous vehicles under its drivered and driverless testing permits to report annually each instance in which a safety driver took control, under 13 CCR § 227.56. In the year to 30 November 2025 vehicles with testing permits “collectively covered more than 9 million miles” on public roads, and the DMV itself cautions that the reports “are not designed for comparative analysis across companies”.¹¹ The regime is imperfect, and the DMV has proposed replacing disengagement reporting with new metrics. But the principle stands: autonomy was extended to systems whose interventions were counted and disclosed, along with the miles driven.

The three precedents differ in detail and agree on the shape. Nobody certified “the aircraft software”; they certified each function to a level. Nobody cleared “medical AI”; they cleared 1,430 devices by the end of 2025, each for its intended use. Nobody permitted “self-driving”; they issued testing permits and demanded the disengagement log. Compliance automation has no equivalent yet. The LegalBench result above, 90 on one category of legal reasoning and 59 on another from the same model, is the reason it needs one.

Evidence that runs both ways

The second thing the precedents share is a record that can be read in either direction: from the requirement to the control that meets it, and from the control back to the requirement that justifies it. Financial regulation already demands this of people, and auditing standards define what counts.

Under PCAOB Auditing Standard 2201, which governs the audit of internal control required by Section 404 of the Sarbanes-Oxley Act, the auditor tests “whether the control is operating as designed” and whether the person performing it “possesses the necessary authority and competence to perform the control effectively”; “performing walkthroughs will frequently be the most effective way” to understand the flow of transactions, following a transaction from origination to the financial records; and a material weakness is “a deficiency, or a combination of deficiencies, in internal control over financial reporting” such that there is “a reasonable possibility that a material misstatement” will go uncaught.¹² Note the order of proof. The auditor does not take the control’s word for it, and the test includes the competence of whoever performs it. When the performer is software, the Stanford hallucination rates are the competence evidence, and they are not yet good enough to pass unexamined.

The FCA’s record-keeping rule puts the direction of evidence in the text. SYSC 9.1.1R requires a firm other than a common platform firm to “arrange for orderly records to be kept of its business and internal organisation” that are “sufficient to enable the FCA to monitor the firm’s compliance with the requirements under the regulatory system”, and SYSC 9.1.2R requires a common platform firm to keep the records of its MiFID business “for a period of at least five years”.¹³ The regulator starts from the requirement and asks for the record; the firm starts from the record and must be able to name the requirement. The AI Act adds the machine’s own trail: logs kept for at least six months, overseers able to “correctly interpret the high-risk AI system’s output”.⁷ An automated compliance system that cannot show, for every control it operates, the clause it serves and the evidence it produced, fails these tests the way an undocumented spreadsheet would.

Where the data sits

The fourth condition concerns location, and two of the texts that govern it, the GDPR and DORA, are EU regulations rather than guidance. A compliance system reads sensitive material: client records, trade data, suspicious activity files, internal investigations. Where that material is processed is regulated.

The GDPR permits a transfer of personal data to a third country only where Chapter V’s conditions “are complied with by the controller and processor”, so that the level of protection “is not undermined”; absent an adequacy decision, Article 46 allows it “only if the controller or processor has provided appropriate safeguards, and on condition that enforceable data subject rights and effective legal remedies for data subjects are available”.¹⁴ DORA, the EU’s digital resilience regulation for finance, requires every ICT contract to state “the locations, namely the regions or countries, where the contracted or subcontracted functions” are provided “and where data is to be processed, including the storage location”, with the provider bound “to notify the financial entity in advance if it envisages changing such locations”; for ICT services supporting critical or important functions the firm must hold “unrestricted rights of access, inspection and audit” and “exit strategies”, and in all cases it “remain[s] fully responsible for compliance with, and the discharge of, all obligations”.¹⁵

The UK rule is older and blunter. Under SYSC 8.1.6R a firm that outsources critical or important operational functions “remains fully responsible for discharging all of its obligations under the regulatory system”, and for a common platform firm SYSC 8.1.8R(9) requires that “the firm, its auditors, the FCA and any other relevant competent authority must have effective access to data” and to the provider’s premises.¹⁶ The European Banking Authority’s 2019 outsourcing guidelines state that “the outsourcing of functions cannot result in the delegation of the management body’s responsibilities” and require the firm’s outsourcing register to record “the location (i.e. country or region) of the data” for every arrangement.¹⁶

None of these texts forbids a cloud or a vendor. Together they make the firm answerable for where its data is, able to say so to the regulator, and able to leave. A compliance tool that sends the firm’s records to a model it does not control, in a country it did not choose, under terms it cannot audit, can put the firm in breach of rules that predate the tool. One way to meet the location requirements by design is for the data not to leave the firm’s own environment.

The four conditions

Put the evidence together and a standard falls out of it. Nothing in it is new; each line is already a rule, a supervisory expectation or a measured precedent. The proposal here is to apply all four to a single class of software before it is allowed to act without a person in the loop.

1. A named accountable person. Every automated compliance function is assigned to an identified individual who holds the regulatory responsibility for it: a Senior Manager under the SM&CR, a member of the management body under MiFID II and ESMA’s statement, the natural person with “competence, training and authority” under Article 26 of the AI Act.⁶⁷ The machine does not appear on the Statement of Responsibilities; the person does.
2. Measured accuracy per task, with a published threshold before a task is automated. Each distinct task the system performs is tested on a held-out set the firm controls, and the result is recorded alongside the threshold the accountable person set in advance, as the US banking agencies’ model-risk guidance includes outcomes analysis in validation and the FDA reviews evidence per intended use.⁷¹⁰ A task that has not met its threshold is performed by the machine only with a person reviewing every output. The LegalBench spread, 59 to 90 for one model across six categories, is why the threshold is per task and not per product.³
3. Traceable evidence in both directions. From every rule to the controls that serve it, and from every control, decision and record back to the rule, with the system’s own logs retained, in the form AS 2201, SYSC 9 and Article 26(6) already describe.⁷¹²¹³ A regulator starting from a clause and an auditor starting from a transaction arrive at the same record.
4. Data in the firm’s own environment. Client, transaction and investigative data stays within infrastructure the firm controls and can locate, audit and exit, meeting the data-location and data-access requirements of GDPR Chapter V, DORA Article 30, SYSC 8 and the EBA register by construction rather than by contract.¹⁴¹⁵¹⁶

This is the standard the letter proposes for any automated compliance system, and it applies to the one Reg42 is building. On Reg42’s own account the four conditions are the design intent: a named owner for every automated duty, a per-task accuracy threshold published before that task runs unattended, two-way lineage from clause to control and back, and deployment inside the customer’s environment. None of that is demonstrated yet, and this article has no measurement of it. The point of writing the conditions down is that they can be checked. A vendor that cannot show a threshold has not earned the task; a firm that cannot name the person has not delegated it, whatever the contract says.

The evidence also sets limits on the argument. Aviation has had a software standard since 1982¹¹ and still records 1.32 accidents per million flights⁸, so proof before autonomy is a discipline, not a guarantee. The Vals result, machines at 80% against lawyers at 71%, shows that a measured threshold can also expose the humans, which cuts against anyone who wants a person in the loop for its own sake.⁴ And the Stanford studies measured legal research, not compliance mapping; none of the studies cited here reports the accuracy of a model that reads a rulebook and writes an obligation, and this article does not measure it either. That is the number the industry owes itself before the first firm lets the map draw itself.

Notes and sources

1. Matthew Dahl, Varun Magesh, Mirac Suzgun and Daniel E. Ho, Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models, Journal of Legal Analysis 16, no. 1 (2024): 64 to 93, doi 10.1093/jla/laae003; arXiv 2401.01301, v2 of 21 June 2024. arXiv abstract: hallucinations, “textual output that is not consistent with legal facts”, occur “between 58% of the time with ChatGPT 4 and 88% with Llama 2” when the models are asked “specific, verifiable questions about random federal court cases”. v2 PDF, Figure 1 (all reference-based tasks pooled): GPT-4 0.58, GPT-3.5 0.69, PaLM 2 0.72, Llama 2 0.88. Figure 8 (all contra-factual tasks pooled): GPT-4 0.69, GPT-3.5 0.52, PaLM 2 0.99, Llama 2 0.02; Llama 2’s statements that a case or justice does not exist are recorded “as successful rejections of the premise”. The authors “execute more than 800,000 queries”.
2. Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning and Daniel E. Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies 22 (2025) 216 to 242, doi 10.1111/jels.12413; preprint 30 May 2024; author-hosted copy. Abstract: the tools “each hallucinate between 17% and 33% of the time” and “the providers’ claims are overstated”. Text: Lexis+ AI “answering 65% of our queries accurately”; Westlaw AI-Assisted Research “accurate 42% of the time” and “one-third of its responses contained a hallucination”; “Over 1 in 6 of our queries caused Lexis+ AI and Ask Practical Law AI to respond with misleading or false information”; Ask Practical Law AI “is a more limited product”, and its low responsiveness “can be explained by its more limited universe of documents”. Figure 4 and Table C1 (counts out of 202): accurate 65%, 42% and 20%; incomplete (refusals or ungrounded responses) 18%, 25% and 63%; hallucinated 17%, 33% and 17%, for Lexis+ AI, Westlaw AI-AR and Ask Practical Law AI. The results text prints 41% for Westlaw and 19% and 62% for Ask Practical Law AI; footnote 17 states that a revalidation raised Ask Practical Law AI’s accuracy “from 19% to 20%”. 202 preregistered queries, run 22 March to 22 April 2024 (Westlaw: 23 to 27 May 2024). No GPT-4 percentage from this study is used here.
3. Neel Guha et al., LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models, 2023. 162 tasks from 36 sources in six categories; abstract: tasks “designed and hand-crafted by legal professionals”. Table 2 (average performance): issue-spotting GPT-4 82.9 / GPT-3.5 60.9; rule-recall 59.2 / 46.3; rule-conclusion 89.9 / 78.0; interpretation 75.2 / 72.6; rhetorical-understanding 79.4 / 66.7. Table 3: rule-application correctness 82.2 / 58.5. The authors note that LegalBench “only considers English language tasks, is skewed towards certain jurisdictions (American law)” and that the results are “a lower-bound on performance, as better prompts may elicit higher scores”.
4. Vals AI, VLAIR Legal Research report (October 2025, page updated 14 October 2025) and application reports for Alexi (80%), Counsel Stack (81%, responses to 196 of 200 questions) and Midpage (78%), each updated 15 October 2025: accuracy on 200 US legal research questions, lawyer baseline 71%, average of all participants 80%. The report: “The evaluated products were Alexi, Counsel Stack, Midpage and ChatGPT” (ChatGPT 80% accuracy); the lawyer baseline came from lawyers supplied by a US law firm; responses were collected over the first three weeks of July 2025; each vendor’s product is evaluated “with their explicit approval”; and “Vals AI has a customer relationship with one or more of the participants”.
5. Harvey, Introducing BigLaw Bench, 29 August 2024: Harvey’s models are described as “producing 74% of a final, expert lawyer-quality work product”; the answer score asks “What % of a lawyer-quality work product does the model complete for the user?” and “Negative criteria, such as hallucinations, were assigned negative scores.” A vendor-published benchmark on the vendor’s own tasks.
6. Financial Conduct Authority, AI Update, April 2024, paragraphs 3.40 and 3.41: the SM&CR “emphasises senior management accountability and is relevant to the safe and responsible use of AI”; dual-regulated firms and solo-regulated Enhanced SM&CR firms must ensure that one or more Senior Managers “have overall responsibility for each of the activities, business areas, and management functions of the firm”, so that “any use of AI in relation to an activity, business area, or management function of a firm would fall within the scope of a SMF manager’s responsibilities”; all Senior Managers in SM&CR firms must have a Statement of Responsibilities and “take reasonable steps to ensure that the business of the firm, for which they are responsible, is effectively controlled”. Foreword: “the FCA is a technology-agnostic, principles-based and outcomes-focused regulator”. Bank of England and FCA, FS2/23, Artificial Intelligence and Machine Learning, 26 October 2023: 54 responses to DP5/22; “Most respondents did not think that creating a new Prescribed Responsibility (PR) for AI to be allocated to a Senior Management Function (SMF) would be helpful for enhancing effective governance of AI” (4.17).
7. European Securities and Markets Authority, Public Statement on the use of Artificial Intelligence in the provision of retail investment services, ESMA35-335435667-5924, 30 May 2024, paragraphs 2, 10 and 23 to 24. FINRA, Regulatory Notice 24-09, 27 June 2024. US Securities and Exchange Commission, Division of Examinations, Examination Priorities, Fiscal Year 2025, section VII.B. Regulation (EU) 2024/1689 (the AI Act), Official Journal text on EUR-Lex (OJ L, 2024/1689, 12 July 2024), Article 14(1) and (4) and Article 26(2), (5) and (6). Board of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation and Office of the Comptroller of the Currency, SR 26-2, Revised Guidance on Model Risk Management, 17 April 2026, which “supersedes and replaces SR letter 11-7” (issued 4 April 2011) and “reflects supervisory experience and industry feedback accumulated over the past fifteen years”; attachment, sections II, III and V and footnote 3.
8. IATA, Safety Report 2025, full year, and press release of 9 March 2026: all-accident rate 1.32 per million flights in 2025 (1.42 in 2024; 2021 to 2025 average 1.27; 2016 to 2025: 1.61, 1.06, 1.45, 1.15, 1.51, 1.13, 1.30, 1.16, 1.42, 1.32); 51 accidents, 8 fatal, 394 on-board fatalities, 38.7 million flights; fatality risk 0.17 (2016 to 2025: 0.18, 0.09, 0.12, 0.09, 0.09, 0.22, 0.11, 0.03, 0.06, 0.17), which “measures the exposure of a passenger or crew to a catastrophic accident with no survivors”. The report notes that the accident definition was revised in 2023 and again in 2025 and that “The dataset presented in this report does not exactly match earlier editions due to the revised definition of accident criteria and improved sectors and accident information”; IATA’s February 2025 release had printed 1.13 for 2024, before those revisions.
9. Youn-Soo Lee and Bo-Young Youn, Clinical Specialty Expansion of AI-Enabled and Machine Learning-Enabled Medical Devices Authorized by the US Food and Drug Administration From 1995 to 2025: Longitudinal Content Analysis, Journal of Medical Internet Research 2026;28:e103040. 1,430 devices with final decisions on or before 31 December 2025 (list downloaded 1 March 2026): 42 in 1995 to 2015, 190 in 2016 to 2019, 406 in 2020 to 2022, 792 in 2023 to 2025, of which 331 in 2025. Per-year averages computed for this article by dividing each total by the years in the period (21, 4, 3, 3).
10. US Food and Drug Administration, Artificial Intelligence-Enabled Medical Devices, content current as of 22 September 2026: “The FDA has authorized over 1,600 AI-enabled medical devices for marketing in the United States as of September 2026”; devices are reviewed “through an appropriate premarket pathway, such as” 510(k) clearance, De Novo classification or premarket approval. The list page (content current as of 4 September 2026) states that the listed devices have had “a focused review of the device’s overall safety and effectiveness, which includes an evaluation of study appropriateness for the device’s intended use and technological characteristics”, that “The list is not a comprehensive resource of AI-enabled medical devices” and that devices were identified “primarily based on the use of AI-related terms in the summary descriptions” of their authorisation documents.
11. Federal Aviation Administration, Advisory Circular 20-115D, Airborne Software Development Assurance Using EUROCAE ED-12( ) and RTCA DO-178( ), 21 July 2017, paragraph 6: “ED-12C/DO-178C is an acceptable means of compliance for the software aspects of type certification or TSO authorization”, and 6.a: applicants “should satisfy all of the objectives associated with the software level assigned to the software”; paragraph 9.b(2): “The system safety process assigns the minimum development assurance level based on the severity classifications of failure conditions for a given function”; paragraph 11.c lists RTCA DO-178 as dated January 1982. California Department of Motor Vehicles, press release of 20 February 2026 on the annual autonomous vehicle disengagement reports: since 2014 California regulations have required manufacturers operating autonomous vehicles under drivered and driverless testing permits to report annually instances where safety drivers took control; vehicles with testing permits “collectively covered more than 9 million miles” on public roads from 1 December 2024 to 30 November 2025; the reports “are not designed for comparative analysis across companies”; and new regulations “propose to replace disengagement reporting requirements previously set forth in 13CCR § 227.56 with updated metrics”.
12. Public Company Accounting Oversight Board, AS 2201: An Audit of Internal Control Over Financial Reporting That Is Integrated with An Audit of Financial Statements, paragraphs .37, .44 and .A7. Its footnote 2 states that it “is the standard on attestation engagements referred to in Section 404(b) of the Act”, the Sarbanes-Oxley Act of 2002.
13. FCA Handbook, SYSC 9.1, as in force on 20 October 2026: SYSC 9.1.1R requires a firm other than a common platform firm to “arrange for orderly records to be kept of its business and internal organisation” that are “sufficient to enable the FCA to monitor the firm’s compliance with the requirements under the regulatory system”; SYSC 9.1.1A(2)R (common platform firms) requires records “sufficient to enable the FCA to fulfil its supervisory tasks”; SYSC 9.1.2R requires a common platform firm to retain the records of its MiFID business “for a period of at least five years”.
14. Regulation (EU) 2016/679 (GDPR), Official Journal text on EUR-Lex, Articles 44 and 46. Article 44: a transfer to a third country “shall take place only if” the conditions of Chapter V “are complied with by the controller and processor”, so that the level of protection “is not undermined”. Article 46(1): “In the absence of a decision pursuant to Article 45(3), a controller or processor may transfer personal data to a third country or an international organisation only if the controller or processor has provided appropriate safeguards, and on condition that enforceable data subject rights and effective legal remedies for data subjects are available.”
15. Regulation (EU) 2022/2554 (DORA), Official Journal text on EUR-Lex (OJ L 333, 27 December 2022), Articles 28 and 30. Article 28(1)(a): financial entities “remain fully responsible for compliance with, and the discharge of, all obligations under this Regulation”; Article 28(8): “For ICT services supporting critical or important functions, financial entities shall put in place exit strategies.” Article 30(2)(b): contracts must state “the locations, namely the regions or countries, where the contracted or subcontracted functions” are provided “and where data is to be processed, including the storage location”, with “the requirement for the ICT third-party service provider to notify the financial entity in advance if it envisages changing such locations”; Article 30(3)(e)(i), for ICT services supporting critical or important functions: “unrestricted rights of access, inspection and audit by the financial entity, or an appointed third party, and by the competent authority”; Article 30(3)(f): “exit strategies”.
16. FCA Handbook, SYSC 8.1, as in force on 20 October 2026: under SYSC 8.1.6R a firm other than a common platform firm that outsources critical or important operational functions or any relevant services and activities “remains fully responsible for discharging all of its obligations under the regulatory system” (SYSC 8.1.6-AR applies the same principle to a common platform firm under the UK law on markets in financial instruments); under SYSC 8.1.8R(9), for a common platform firm, “the firm, its auditors, the FCA and any other relevant competent authority must have effective access to data related to the outsourced activities, as well as to the business premises of the service provider, where necessary for the purpose of effective oversight”. European Banking Authority, Guidelines on outsourcing arrangements, EBA/GL/2019/02, 25 February 2019: paragraph 35, “The outsourcing of functions cannot result in the delegation of the management body’s responsibilities”; paragraph 54(f), the outsourcing register must record, for all existing outsourcing arrangements, “the country or countries where the service is to be performed, including the location (i.e. country or region) of the data”.