Source-Credit Fabrication on CCP override framework disclosure. A frontier AI model, asked what CCPs must publicly disclose about their margin model override framework under the CPMI-IOSCO Initial Margin Disclosure Consultation, returned a confident three-part enumeration that does not appear in the consultative document and framed an expectation as a mandatory obligation. The error is the kind that survives a quick read of an opinion letter and only surfaces when the client is examined against the actual regulator text.

The pattern in one line

A frontier AI assistant, asked a precise question about a discretionary disclosure category in the CPMI-IOSCO Initial Margin Disclosure Consultation, returns a confident, plausibly drafted three-element list that is not in the consultation text, and frames the obligation as a mandatory requirement when the consultation uses an expectation standard.

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 Panel posed a Specialist Panel application-style question to two frontier AI subjects, framing the request the way a junior associate would type it into an AI assistant when preparing a client memo on CCP disclosure scope.

The Panel then bound the model output against the verbatim consultation text on CCP override framework disclosure held as primary substrate, and against the broader d232 cover note on CCP margin model overrides.

The Panel does not give the AI subjects the substrate. The Panel evaluates only the AI subject's own output against the substrate the Panel holds. That asymmetry is what makes the finding usable: it shows how the model behaves when the lawyer or junior associate is not yet looking at the regulator text.

What the models got wrong

Claude Sonnet 4.6, web search on, asked what specific information CCPs must publicly disclose about their margin model override framework under the 2026 consultation on initial margin transparency, returned a structured three-part answer: (1) the instances or circumstances where overrides may be warranted; (2) the key decision-makers authorised to exercise override discretion; and (3) the permissible types of adjustments that can be made.

Each of those three categories sounds like the kind of category a CCP override framework disclosure would naturally cover. None of them appears in the consultation text. The consultation states: "CCPs should publicly disclose relevant information on their override framework." The specific content of "relevant information" is left open for comment. The model did not return "should". It returned "must". It did not say "the consultation leaves the specific content open". It enumerated three categories with the confidence of a settled standard.

The model also cited a secondary commentary URL as the source for the three-part enumeration, rather than the primary BIS d232 cover note. A lawyer who follows the citation gets a third-party reading of the consultation, not the consultation. A lawyer who does not follow the citation gets a sentence that reads as if it were drawn from the BIS document.

Citation: RLB-H-INT-BIS-CPMI-IOSCO-INITIAL-MARGIN-DISCLOSURE-CONSULT-2026-Q005-Sonnet46.

Why this matters for lawyers

A board memo, opinion letter, or comment-letter submission that embeds the three-part enumeration creates a deliverable in which the client's disclosure framework is structured around requirements the regulator did not set. The downstream consequences for a CCP client are concrete: the client's disclosure template is drafted to a specification the consultation does not contain; the audit committee signs off on adequacy against that specification; the regulatory dialogue is conducted on the basis that those three categories are the test of sufficiency.

If the client is a clearing member or a prime broker rather than the CCP itself, the same memo flows into the counterparty due diligence regime. The clearing member or prime broker may flag CCPs as deficient against the fabricated standard, trigger remediation correspondence with the CCP, or raise the issue in its own regulatory mapping. None of that activity has a regulatory basis in the consultation text.

The professional indemnity exposure is the exposure that an opinion or memo, signed by a partner, asserts a position on the regulator's expectation that the regulator's text does not support. The exposure is not theoretical: the consultation is open for comment, and the comments themselves will scrutinise what the consultation does and does not say. A comment-letter submission that asserts the three-part enumeration on behalf of a client will be on the public record against the consultation it purports to describe.

The regulator's actual position

The consultative document states, verbatim: "CCPs should publicly disclose relevant information on their override framework." The cover note for d232 sets the override framework disclosure as one of several areas in which CCPs are expected to provide relevant information; it does not enumerate the contents of that disclosure, and it does not frame the obligation as a hard requirement. The use of "should" is the CPMI-IOSCO Secretariat's standard formulation for an expectation that anchors supervisory dialogue without prescribing line-item content.

The decision to leave the specific content open for comment is itself a substantive feature of the consultation. The Secretariat has not pre-judged what "relevant information" comprises. That openness is what the comment process is designed to resolve.

What this tells us about AI for lawyers

The failure class on this finding is what the Panel labels Source-Credit Fabrication: the model attaches a confident, structured factual claim to a regulator-issued source that does not contain the claim, and supports the attribution with a secondary commentary URL. The failure is not a paraphrase drift. It is a generation of facts that read as if they were drawn from the regulator's text.

For lawyers, the operational signal is that any AI output reporting "the regulator requires (i), (ii), and (iii)" must be verified against the cited regulator text, and the citation itself must be verified against the primary source. A secondary commentary URL is not a source. A confident three-part enumeration on a consultation that is open for comment is the highest-risk class of AI output a lawyer will see, because it is the class that looks most like a settled standard.

The second signal is that AI output framing a "should" expectation as a "must" requirement is the most common drift class on consultation-stage regulatory text. The drift is small in word count and large in legal effect. The lawyer's review pass must catch it every time.

What the RLB Specialist Panel is doing about it

The Panel issues the finding bound to the verbatim consultation text held as primary substrate, with the citation ID assigned for cross-reference by AI labs, by frontier model evaluation teams, and by enterprise legal and compliance teams that subscribe to RLB briefings. The Panel does not assess the model's overall accuracy. The Panel assesses the specific output against the specific regulator text on the specific question a practitioner is likely to ask.

For AI labs, the finding is available for incorporation into training-data and evaluation pipelines under the partnership track. For lawyers and law firms, the finding is available as a worked example of the verification discipline required when AI is used in regulatory-mapping work product.

Citation: RLB-H-INT-BIS-CPMI-IOSCO-INITIAL-MARGIN-DISCLOSURE-CONSULT-2026-Q005-Sonnet46.

What lawyers teams should do

Action items for a law firm regulatory practice using AI on consultation-stage CCP disclosure work:

• Verify every AI-returned enumeration of regulator obligations against the primary regulator text: a confident three-part list is the highest-risk class of AI output on consultation-stage documents.
• Treat any AI output framing a consultation as imposing a "must" requirement as a presumptive drift: consultations almost always use expectation language, and the drift from "should" to "must" is the most common AI drift on this document class.
• Reject secondary commentary URLs as citation sources for opinion-letter content: if the AI cites a third-party commentary rather than the primary BIS text, treat the underlying factual claim as unverified.
• Build a pre-circulation check into the partner sign-off workflow that requires the drafter to identify, against the primary regulator text, the verbatim source for each enumerated obligation in the memo.
• Where a client is a CCP, a clearing member, or a prime broker, scope the engagement letter to capture that AI-assisted drafts are verified against primary regulator text before partner sign-off, and document the verification step in the file.
• Use the RLB Specialist Panel's published findings as a working library of AI failure modes on regulator text, and brief associates on the failure-class pattern when onboarding to the practice.

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 Consultative Report d232, Streamlining Variation Margin Disclosure (2026) · Substrate documents: p_03_ANNEX_Override_framework___public_disclosure_r_d232_covernote.pdf · CPMI portal: bis.org/cpmi

Citation IDs referenced:

• RLB-H-INT-BIS-CPMI-IOSCO-INITIAL-MARGIN-DISCLOSURE-CONSULT-2026-Q005-Sonnet46