Inverted modality: frontier AI mis-classifies CPMI-IOSCO d226 effective practices as supervisory or mandatory obligations.

The document itself records that it provides "examples of how standards set out in the CPMI-IOSCO Principles for financial market infrastructures, as supplemented by the relevant guidance, can be met." A frontier AI model tested by the RLB Specialist Panel produced a complete board-ready compliance obligations memo that classified every one of the eight effective practices as either a supervisory expectation in its own right or as overlapping with mandatory national rules.

The pattern in one line

A frontier AI model, asked by a lawyers user to classify each of the eight effective practices in CPMI-IOSCO d226 as mandatory, supervisory, or voluntary, returned a full memo that converted a voluntary BIS publication into a supervisory baseline. The document's own stated purpose paragraph records the opposite. The model's commitment is the kind of confident, citable text that a lawyers user would paste into a deliverable before verification against the source.

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. For this finding, the Specialist Panel ran a Specialist Panel application-style question that placed the model in the role of a CCP General Counsel preparing a compliance obligations memo for the board's Audit and Risk Committee.

The question required the model to classify each of the eight effective practices in d226 as either (A) MANDATORY REQUIREMENT with enforcement consequences, (B) SUPERVISORY EXPECTATION that regulators will test against, or (C) VOLUNTARY GUIDANCE with no binding legal force, to cite exact language from the document for each classification, and to identify the enforcement mechanism if any. Brevity and generic-language hedging were treated as non-response. The model's full output was then compared against the d226 final report, where the document's own stated purpose paragraph records the binding-force position.

What the models got wrong

Claude Opus 4.7, queried with web search enabled, produced a complete compliance obligations memo addressed from "General Counsel" to the "Board Audit and Risk Committee." The memo opened with a threshold classification: "d226 is described as 'examples of how standards set out in the CPMI-IOSCO Principles for Financial Market Infrastructures (PFMI), as supplemented by the relevant guidance, can be met.' The document is therefore properly classified as (C) VOLUNTARY GUIDANCE in its own right, but with a strong gravitational pull into (B) SUPERVISORY EXPECTATION because the underlying PFMIs are the de facto binding standard against which CCPs are supervised."

The memo then classified each of the eight practices. Practice 1 (scheduled intraday VM calls): (B) Supervisory expectation. Practice 2 (ad hoc intraday VM calls): (B) Supervisory expectation. Practice 3 (use of excess collateral to meet VM): (C) Voluntary guidance, trending (B) where the CCP rulebook already references it. Practice 4 (pass-through of VM by CCPs): (B) Supervisory expectation. Practice 5 (acceptable forms and currency of VM): (B) Supervisory expectation, with (A) elements. Practice 6 (pass-through of VM between clearing members and clients): (B) Supervisory expectation. Practice 7 (CCP transparency on VM processes): (B) Supervisory expectation, with (A) overlap.

Practice 8 (clearing-member transparency to clients): (B) Supervisory expectation.

The bottom-line paragraph of the memo then advised the Committee that "all eight are properly treated as (B) SUPERVISORY EXPECTATIONS" and that the Committee "should plan implementation on the assumption that material non-adherence will draw supervisory findings." The d226 final report does not support that conclusion. The Specialist Panel records the failure under the misstated-rule category. The full finding analysis, the verbatim regulator text, and the AI's complete answer are at RLB-H-INT-BIS-CPMI-CPMI-IOSCO-VARIATION-MARGIN-CCPs-2025-Q004-Opus47.

Why this matters for Lawyers

Partner-level memos and board briefings that classify d226 practices as supervisory expectations or mandatory obligations create disclosure exposure to clients, audit committees, and counterparties who rely on the legal characterisation to size implementation budgets, draft rulebook amendments, and respond to supervisory dialogue. The pattern is particularly difficult to catch because the AI's output is internally coherent: the threshold classification correctly identifies d226 as voluntary, but then immediately overrides that identification with a supervisory-expectation gloss that runs through the entire memo. A reviewer who reads only the bottom-line paragraph or only the practice-by-practice table sees a confident supervisory characterisation.

A reviewer who reads only the threshold paragraph sees a voluntary characterisation. Both readers may sign off on the memo. The inconsistency is the failure.

The regulator's actual position

The d226 final report sets out its own stated purpose in unambiguous terms. The document provides "examples of how standards set out in the CPMI-IOSCO Principles for financial market infrastructures, as supplemented by the relevant guidance, can be met." The document does not create new supervisory obligations, does not extend the PFMIs, and does not impose new mandatory rules on CCPs or clearing members. The underlying PFMI Principles remain the binding standard against which CCPs are supervised, but the d226 publication itself is voluntary illustration of one way the PFMI standards can be met.

The Specialist Panel records this verbatim text as the primary source against which the AI's commitment is compared. The R-folder reference is R6-FINAL_REPORT-00001.

What this tells us about AI for Lawyers

Across the d226 evaluation, the failure surfaced when the AI was placed under deliverable pressure: the prompt asked for a full compliance memo with specific classifications and cited language. Under that pressure, the model committed to a supervisory baseline even where it had correctly identified the document's voluntary status in its own threshold paragraph. That suggests a generation pattern where the model resolves uncertainty about the binding force of an international standard-setter publication by defaulting to the more demanding characterisation, on the implicit theory that supervisors will examine adherence.

For Lawyers, the lesson is that AI commitments on binding force in deliverables targeted at boards, audit committees, or regulators should be treated as flags for verification against the document's own purpose paragraph, not as authoritative legal characterisations.

What the RLB Specialist Panel is doing about it

The RLB Specialist Panel documents this finding under Citation ID RLB-H-INT-BIS-CPMI-CPMI-IOSCO-VARIATION-MARGIN-CCPs-2025-Q004-Opus47 and binds it to verbatim regulator-issued source text from the d226 final report. The finding is available for inspection at /regulators/j1/INT/BIS-CPMI-INT-001/CPMI-IOSCO-VARIATION-MARGIN-CCPs-2025/ai-labs/finding/INT-BIS-CPMI-INT-001-CPMI-IOSCO-VARIATION-MARGIN-CCPs-2025-v1-004--opus-47-websearch/. The Panel is interested in partnership with frontier AI labs and with practitioners who rely on AI for international standard-setter material. AI labs receive the per-finding analysis with training-data and post-training-logic implications under the AI Labs whitepaper at /regulators/j1/INT/BIS-CPMI-INT-001/CPMI-IOSCO-VARIATION-MARGIN-CCPs-2025/ai-labs/. Practitioners receive the audience-specific case study at the cell summary page for their profession or sector and department.

What Lawyers teams should do

• Treat any AI-generated classification of d226 effective practices as binding, supervisory, or mandatory as a flag for direct verification against the document's own "examples of how standards … can be met" framing.
• When advising on d226 implementation, separate the document's own legal status (voluntary illustration of PFMI compliance pathways) from the binding force of the PFMI Principles themselves and from any national rulebook that incorporates a specific practice.
• Maintain a verified extract of d226's stated purpose paragraph in the firm's regulatory text repository so junior associates can cross-check AI-drafted memos against the regulator's own characterisation.
• When clients ask for jurisdictional treatment of d226, give a separate analysis for each national supervisor's stance rather than treating the BIS publication as a global mandate.
• Flag d226-related AI outputs for senior partner review before any client-circulated work product references binding force.

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: d226 — Streamlining variation margin in centrally cleared markets — examples of effective practices (January 2025). R-folder reference: R6-FINAL_REPORT-00001. BIS portal: bis.org/cpmi.

Citation IDs referenced:

• RLB-H-INT-BIS-CPMI-CPMI-IOSCO-VARIATION-MARGIN-CCPs-2025-Q004-Opus47