Briefings Blog

Two frontier AI models, Two frontier AI models, each running with web search, produced structurally correct-looking operational guidance on CFTC Regulation 1.44 that contradicts the regulation’s text. The Specialist Panel tested each model on the same operational question: map the currency deadline tiers for an FCM margin operations team. Both failed on the regulatory detail that determines how early or late an FCM must collect margin.

Claude Opus 4.7 collapsed the regulation’s three-tier structure into two tiers. Regulation 1.44(f) distinguishes: USD and Canadian dollars (same-day); Appendix A currencies, AUD, CNY, HKD, HUF, ILS, JPY, NZD, SGD, TRY, ZAR, (end of the second business day, T+2); and all other fiat currencies (end of the next business day, T+1). The model assigned Appendix A currencies a T+1 deadline and all other non-USD currencies same-day, compressing an intermediate tier out of existence.

An FCM margin team following this guidance would call Appendix A margin one full business day earlier than required, and apply same-day urgency to currencies the regulation permits T+1 collection on.

Claude Sonnet 4.6 introduced a deadline that does not exist in the regulation: “12:00 p.m. ET on the FIRST U.S. business day.” Regulation 1.44(f)(3) sets only “end of the business day after the day on which the margin call is issued”, no intraday time. A system parameter or operations procedure built on this output would implement a self-imposed noon cutoff with no regulatory basis, and the documentation trail would cite a specific time the regulation does not contain.

Both outputs were formatted, internally consistent, and produced without any caveat that source verification was needed. That is the failure mode these findings document.

Two frontier AI models, Two frontier AI models, both running with web search, generated confident, structurally correct-looking guidance on the CFTC's December 2025 swap dealer business conduct package and its January 2026 correction notice that misnames a staff letter that does not exist, fails to identify the appendix at the centre of the correction, and inverts the scope of a requirement the rule actually preserved.

The RegLeg Brief Specialist Panel tested each model on four separate operational questions a U.S. swap dealer compliance team would put to an AI assistant when briefing a partner-level client: which appendix the January 2026 correction notice restored, which staff letter governs intended-to-be-cleared (ITBC) swap documentation, and what the Pre-Trade Mid-Market Mark (PTMMM) elimination actually covers. On the appendix question, both models were non-responsive, neither named Appendix A to Subpart H of Part 23, the guidance document on §§ 23.434 and 23.440 that the correction was issued to preserve.

On the ITBC swaps question, Opus 4.7 fabricated a "CFTC Staff Letter 25-49" whose substantive content does not exist in the regulatory record. The operative letter is CFTC Staff Letter 23-01, which superseded Letter 13-70. On the PTMMM scope question, Opus 4.7 stated the elimination was "product-agnostic across the desk's covered swap book." The CFTC eliminated paragraph (a)(3) of § 23.431 in its entirety, but the disclosure exceptions in § 23.431(c) remain in force for SEF and DCM-initiated trades and anonymous counterparties.

For swap dealers and major swap participants subject to External Business Conduct Standards, the implication is operational. A partner-level advisory drafted on the Opus 4.7 output would cite a non-existent CFTC staff letter to a client. A documentation-process redesign built on the PTMMM answer would lift price and compensation disclosure controls on trades the regulation still subjects to those controls.

The Specialist Panel calls this failure class Amendment-Layer Misattribution and is documenting the verbatim model outputs, regulator excerpts, and methodology in an open-access white paper alongside the immutable Citation IDs RLB-H-US-CFTC-SWAP-DEALER-BUSINESS-CONDUCT-DOCUMENTATION-2025-Q002-Opus47, RLB-H-US-CFTC-SWAP-DEALER-BUSINESS-CONDUCT-DOCUMENTATION-2025-Q002-Sonnet46, RLB-H-US-CFTC-SWAP-DEALER-BUSINESS-CONDUCT-DOCUMENTATION-2025-Q003-Opus47 and RLB-H-US-CFTC-SWAP-DEALER-BUSINESS-CONDUCT-DOCUMENTATION-2025-Q004-Opus47.

Two regulatory queries on MAS Notice 637, the operative risk-based capital adequacy framework for Singapore-incorporated banks, produced confident output from a frontier AI model with web search active that contradicts text on the face of the regulator's own documents. Both findings sit inside a failure class the RegLeg Brief Specialist Panel has been documenting across Singapore, U.S., and international supervisory work, Confident Fabrication of Adjacent Regulatory Instruments, the pattern of frontier AI inventing plausibly named, plausibly numbered regulatory instruments that do not exist, and misrepresenting drafting-aid annotation as substantive regulatory text.

In the first finding, Opus 4.7 was asked whether MAS Notice 637 applies to Singapore-incorporated financial holding companies and whether a separate MAS notice exists for them. The model produced the answer that FHCs are governed by "Notice FHC-N637 (Risk Based Capital Adequacy Requirements for Financial Holding Companies)", a notice that does not exist on the MAS Notices and Directives register. The actual position is that Notice 637 applies, by its own paragraph 1.1, to Reporting Banks, and FHCs are governed by a separate MAS notice under the Financial Holding Companies Act, which the model did not name.

In the second finding, Opus 4.7 was asked what yellow-highlighted passages in the MAS Notice 637 (Amendment) 2024 PDFs signify and whether they will appear in the consolidated Notice. The model described the yellow as drafting visual emphasis, when the cover note on the amendment package states explicitly that the yellow highlighting is annotation describing the change and will not appear in the published untracked Notice. The model's reading inverts the regulator's stated convention.

The findings are published with the immutable RLB Citation IDs RLB-H-SG-MAS-NOTICE-637-CAPITAL-ADEQUACY-BANKS-2025-Q010-Opus47 and RLB-H-SG-MAS-NOTICE-637-CAPITAL-ADEQUACY-BANKS-2025-Q012-Opus47.

Company Secretaries at FMIs and at organisations whose governance documentation references CPMI-IOSCO supervisory output are increasingly using AI to draft board pack methodology notes on Level 3 assessments, prepare counterparty disclosure summaries on regulator engagement, produce committee minute extracts characterising the assessment lifecycle, and validate procedural-fact statements destined for the audit-committee record. The November 2025 CPMI-IOSCO Level 3 assessment of general business risk, recorded under PFMI Principle 15, is the supervisory exercise most directly bearing on this practice area in the current cycle.

As AI tooling enters the drafting layer, the question is no longer whether AI-assisted work product reaches client-facing deliverables; it is whether the work product reaches them with the regulator-text fidelity that Company Secretaries need.

The RLB Specialist Panel tested two frontier AI models on a question set covering the LNAFE quantitative floor, the Basel/CRD equity carve-out condition, and the November 2025 assessment lifecycle. The Panel records 1 finding on this audience-specific cell. The failure pattern in scope: Supervisory-timeline truncation, dropping the validation phase. Questions are prepared by the RLB Specialist Panel based on real practical AI usage in the workflows the respective audience uses AI for. The Panel binds each AI finding to verbatim regulator-issued source text held as primary substrate.

For Company Secretaries the operational consequence is direct. A board pack that records the CPMI-IOSCO Level 3 assessment as having run "during 2023 and 2024" misrepresents a supervisory process by truncating its 2025 findings-sharing and validation phase, and any downstream document that inherits that draft, whether a counterparty memorandum, a peer-review submission, or a regulator engagement summary, will carry the same factual inaccuracy.

PFMI Principle 15 is one of the cleanest primary-source surfaces in the cross-border CCP and CSD universe: a Key Consideration cited in a deliverable is either the right KC or it is not; a quantitative floor is either the regulator's text or it is not; an assessment-period date range is either accurate or it is not. Each is recoverable on a routine line-by-line read.

The audit's 1 finding for this cell carry immutable RLB Citation IDs and are bound to verbatim regulator-issued source text held by the RLB Specialist Panel: RLB-H-INT-BIS-CPMI-IOSCO-PFMI-L3-GENERAL-BUSINESS-RISK-2025-Q005-Sonnet46. The full audit on the November 2025 CPMI-IOSCO Level 3 assessment is published at the PFMI Level 3 General Business Risk hub on RegLegBrief.com.

Accountants advising central counterparty clients on PFMI Principle 15 compliance are increasingly using AI to draft post-assessment remediation work programmes, produce LNAFE sufficiency review notes, validate Basel-versus-LNAFE capital eligibility opinions for CCP audit clients, and prepare partner-level technical memos on the November 2025 CPMI-IOSCO assessment. The November 2025 CPMI-IOSCO Level 3 assessment of general business risk, recorded under PFMI Principle 15, is the supervisory exercise most directly bearing on this practice area in the current cycle.

As AI tooling enters the drafting layer, the question is no longer whether AI-assisted work product reaches client-facing deliverables; it is whether the work product reaches them with the regulator-text fidelity that Accountants need.

The RLB Specialist Panel tested two frontier AI models on a question set covering the LNAFE quantitative floor, the Basel/CRD equity carve-out condition, and the November 2025 assessment lifecycle. The Panel records 2 findings on this audience-specific cell. The failure pattern in scope: Source-text condition replacement with an invented overlay test; Key Consideration mis-attribution of a quantitative threshold. Questions are prepared by the RLB Specialist Panel based on real practical AI usage in the workflows the respective audience uses AI for. The Panel binds each AI finding to verbatim regulator-issued source text held as primary substrate.

For Accountants the operational consequence is direct. An LNAFE sufficiency review note that attributes the six-month operating-expense floor to the wrong Key Consideration, or that advises the client to exclude Basel-grade equity from the LNAFE buffer on the basis of an invented liquidity test, is the kind of document a CCP regulator or its peer reviewer will read line by line during the 2026 cycle.

PFMI Principle 15 is one of the cleanest primary-source surfaces in the cross-border CCP and CSD universe: a Key Consideration cited in a deliverable is either the right KC or it is not; a quantitative floor is either the regulator's text or it is not; an assessment-period date range is either accurate or it is not. Each is recoverable on a routine line-by-line read.

The audit's 2 findings for this cell carry immutable RLB Citation IDs and are bound to verbatim regulator-issued source text held by the RLB Specialist Panel: RLB-H-INT-BIS-CPMI-IOSCO-PFMI-L3-GENERAL-BUSINESS-RISK-2025-Q002-Opus47, RLB-H-INT-BIS-CPMI-IOSCO-PFMI-L3-GENERAL-BUSINESS-RISK-2025-Q003-Sonnet46. The full audit on the November 2025 CPMI-IOSCO Level 3 assessment is published at the PFMI Level 3 General Business Risk hub on RegLegBrief.com.