Source-text condition replacement with an invented overlay test; Key Consideration mis-attribution of a quantitative threshold; and Supervisory-timeline truncation, dropping the validation phase across PFMI Principle 15 KC3 and the November 2025 Level 3 assessment. Two frontier AI subjects tested by the RLB Specialist Panel produced confident, citable answers on the LNAFE quantitative floor, the Basel/CRD equity carve-out condition, and the assessment lifecycle that the regulator's own primary text directly contradicts. The pattern is operationally consequential for Lawyers.

The pattern in one line

Across the audited question set, two frontier AI models tested by the RLB Specialist Panel produced verbatim-looking answers on PFMI Principle 15 Key Consideration 3 and on the November 2025 CPMI-IOSCO Level 3 assessment lifecycle that the regulator-issued source text directly resolves, with 3 findings in scope for the Lawyers cell.

How the RLB Specialist Panel tested this

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. The Panel mapped Principle 15 KC2 and KC3 against the November 2025 executive summary lifecycle recorded by CPMI-IOSCO, and held verbatim regulator-issued source text covering the KC3 quantitative floor, the KC3 Basel/CRD equity carve-out, and the assessment-period record. Two frontier AI subjects were probed across direct-question and application-style framings to capture both straight rule-lookup behaviour and behaviour under realistic client-deliverable drafting load.

Each AI response was bound to the regulator-issued source text it conflicts with. The Specialist Panel records both subject answers and the regulator text in full for each finding, with an immutable RLB Citation ID per finding.

What the models got wrong

Q002, Claude Opus 4.7 (web search on), KC3 Basel/CRD equity carve-out, invented KC4 liquidity overlay. The Specialist Panel asked what capital instruments qualify toward a CCP's minimum LNAFE under PFMI Principle 15, what must be held separately, and whether equity held in compliance with Basel or CRD prudential capital frameworks can be included in the LNAFE calculation. Opus 4.7 answered that Basel-driven capital is not fungible with LNAFE and introduced a fabricated KC4 additivity test, asserting that the firm must demonstrate that the binding requirement is met after netting.

The regulator-issued text in KC3 records the carve-out directly: equity held under international risk-based capital standards can be included where relevant and appropriate to avoid duplicate capital requirements. The Opus answer replaces the duplicate-capital qualifier with an invented liquidity overlay and misallocates the condition to KC4. Citation: RLB-H-INT-BIS-CPMI-IOSCO-PFMI-L3-GENERAL-BUSINESS-RISK-2025-Q002-Opus47.

Q003, Claude Sonnet 4.6 (web search on), KC3 six-month LNAFE floor, misattributed to KC2. Asked, in a derivatives CCP submission framing, what the minimum LNAFE requirement is and what qualifies, Sonnet 4.6 produced a composite "greater of" answer and attributed the quantitative capital floor to KC3 in some passages and to KC2 in others, importing a scenario-analysis leg into the six-month floor that does not appear in the rule. The regulator-issued text records the six-month floor in KC3 only, as a single quantitative minimum.

The Sonnet answer's KC mis-attribution propagates a structural error about the KC architecture into any deliverable that adopts it. Citation: RLB-H-INT-BIS-CPMI-IOSCO-PFMI-L3-GENERAL-BUSINESS-RISK-2025-Q003-Sonnet46.

Q005, Claude Sonnet 4.6 (web search on), assessment timeline, truncated to 2023-2024. Asked when the November 2025 CPMI-IOSCO Level 3 assessment process began, when data was collected, when it formally concluded, how many FMIs participated, and how findings were validated with FMIs, Sonnet 4.6 answered that the assessment work was carried out during 2023 and 2024. The regulator-issued executive summary records the assessment was carried out during 2023-25 by the Implementation Monitoring Standing Group (IMSG) and a team of experts from CPMI and IOSCO member jurisdictions. The Sonnet answer drops the 2025 findings-sharing and validation phase from the recorded lifecycle.

Citation: RLB-H-INT-BIS-CPMI-IOSCO-PFMI-L3-GENERAL-BUSINESS-RISK-2025-Q005-Sonnet46.

Why this matters for Lawyers

Lawyers advising on PFMI Principle 15 are increasingly using AI to draft Principle 15 compliance opinions for central counterparties, produce LNAFE eligibility briefings for liquidity-risk teams, prepare client memos on the November 2025 CPMI-IOSCO Level 3 assessment, and validate Key Consideration cross-references in regulatory submissions and consultation responses. For Lawyers on this practice area, the operational consequence of the failure pattern is direct.

A Principle 15 compliance opinion that misstates the KC3 Basel/CRD equity carve-out condition, or that attributes the six-month LNAFE floor to KC2 rather than KC3, is the kind of memorandum a regulator or counterparty due diligence team will challenge on first read. The 2026 supervisory cycle is recoverable on the regulator's own publications and on the PFMI source text; a deliverable carrying any of the documented failure modes will not survive a careful peer or regulator read.

The risk is amplified when AI-drafted prose is propagated into downstream artefacts, board packs, regulatory submissions, counterparty disclosures, internal training materials, without a primary-source check at each propagation step. Each AI failure recorded in this audit was produced in a confident, fluent register and against a question type that the AI handled without flagging uncertainty, which is precisely the failure mode that survives casual review.

The regulator's actual position

The Bank for International Settlements, through the Committee on Payments and Market Infrastructures and the International Organization of Securities Commissions, records the regulator's position on these three questions directly in the PFMI source text and in the November 2025 Level 3 assessment publication.

On the LNAFE minimum, PFMI Principle 15 Key Consideration 3 states: at a minimum, an FMI should hold liquid net assets funded by equity equal to at least six months of current operating expenses. The text records a single quantitative floor, not a composite. The scenario-analysis obligation referenced by KC2 is a separate prior step in the analysis, not a sizing leg of the KC3 floor.

On the Basel/CRD equity carve-out, Principle 15 KC3 records: equity held under international risk-based capital standards can be included where relevant and appropriate to avoid duplicate capital requirements. The only condition gating inclusion is the duplicate-capital qualifier. There is no liquidity overlay, no KC4 additivity test, and no separate eligibility screen.

On the assessment timeline, the November 2025 executive summary records that the Level 3 exercise was carried out during 2023-25 by the Implementation Monitoring Standing Group and a team of experts from CPMI and IOSCO member jurisdictions, focused on general business risk under Principles 3, 13, 15, and 21. The 2025 phase covers findings-sharing and validation with participating FMIs, and dropping that phase from a procedural representation truncates the supervisory record.

What this tells us about AI for Lawyers

The audited failure modes point to a structural pattern: AI subjects tested on the PFMI Principle 15 substrate produced confident, fluent answers on the highest-stakes structural questions, the quantitative floor, the equity carve-out, and the assessment lifecycle, and did so without flagging uncertainty. For Lawyers the lens is: AI on this surface is not safe for verbatim rule recitation, not safe for KC cross-reference attribution, and not safe for supervisory-timeline statements.

The AI behaviour observed is consistent across both subject models tested: the AI subjects could fluently discuss Principle 15 at the level of general framing, but could not reliably recover the exact KC3 carve-out condition, the exact KC3 six-month floor framing, or the exact 2023-25 assessment period. The implication for AI-assisted work product is that primary-source verification is not optional, it is the workflow. Lawyers carrying this finding into the practice area should treat the three KC3 axes, the quantitative floor, the Basel/CRD carve-out, and the supervisory-cycle date range, as the canonical hot-fact categories on Principle 15.

What the RLB Specialist Panel is doing about it

The RLB Specialist Panel runs ongoing audits across cross-border CCP, CSD, PS, and trade-repository regulation, with primary substrate held verbatim and AI behaviour mapped against the regulator text. The audit findings on this PFMI Level 3 cell are published with immutable Citation IDs, bound to regulator-issued primary substrate, and made available to AI labs and to practitioner audiences as part of an open transparency posture. The Panel is in active dialogue with frontier AI labs on partnership tracks that close the failure loop: from audit finding, to lab-side acknowledgement, to behaviour calibration.

RLB does not characterise these failures as terminal; the Panel records them, binds them to source text, and offers them as a primary input into the AI labs' own calibration and red-team programmes. For practitioner audiences, the audit output also supports the firm's own AI-assisted workflow QC, by surfacing the question types and the failure modes that warrant primary-source verification.

What Lawyers teams should do

The action items below apply to Lawyers working on PFMI Principle 15 and on the November 2025 CPMI-IOSCO Level 3 assessment cycle:

• Treat every Key Consideration citation in an AI draft as a hot fact requiring direct verification against the Principle 15 text.
• Maintain a Principle 15 "KC map" annotated with the LNAFE quantitative floor (KC3), the equity carve-out (KC3), the scenario-analysis obligation (KC2), and the recovery and orderly wind-down plan reference (KC3), to test AI output against.
• Where AI advises on Basel-versus-LNAFE eligibility, confirm that the only condition cited is the duplicate-capital qualifier, not a fabricated liquidity overlay.
• Verify timeline references against the CPMI-IOSCO publication record: the 2025 Level 3 assessment ran 2023-25, not 2023-24.
• Lock a partner-level review of all AI-assisted Principle 15 work product against the primary text before issue.
• Record AI provenance on every internal opinion file so that the firm has a recoverable audit trail of where AI-drafted content was relied on.

Each action item is recoverable against the audit's bound citations and against the regulator-issued primary text. The audit URL is https://reglegbrief.com/regulators/j1/int/bis-cpmi/cpmi-iosco-pfmi-l3-general-business-risk-2025/.

Right of Reply

These findings and associated work have been put up in public with a view of the greater good for the development of a safer AI ecosystem. Any party reading this or any finding on reglegbrief.com may contact us and have an unconditional right of reply; the Specialist Panel will publish any factual correction or contextual response alongside the original finding, with no editorial gatekeeping. Researchers, regulators, and compliance teams with questions on methodology or specific findings can reach the Specialist Panel via the same channel.

Source & Methodology Standards

RegLeg Brief is operated by Verdus Technologies Pte. Ltd. (UEN 201616982R), incorporated in Singapore. The RLB Specialist Panel, with an aggregate of over 60 years of public-policy and industry experience, documents only confirmed hallucination findings, under a methodology that requires a verbatim regulator excerpt for every documented claim. All findings, citation IDs, model outputs, regulator excerpts, and methodology notes are open-access.

Primary source verified: CPMI-IOSCO PFMI Level 3 General Business Risk Assessment (2025) · Substrate documents: p_01_OTHER_d228_Executive_Summary_page_1__full_asse_d228.htm, p_03_NOTICE_d228_Annex_A_reproducing_PFMI_Principle_p251107.htm, p_07_GUIDELINE_PFMI_KC3_existing_standard_vs_FIA_ISDA_r_d162.htm · CPMI portal: bis.org/cpmi

Citation IDs referenced:

• 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
• RLB-H-INT-BIS-CPMI-IOSCO-PFMI-L3-GENERAL-BUSINESS-RISK-2025-Q005-Sonnet46