Deontic Register Failure on the CPMI-IOSCO Initial Margin Consultation 2026

Two frontier AI models running with web search enabled, both tested by the RLB Specialist Panel, produced confidently wrong reconstructions of the CPMI-IOSCO Consultative Report on Updated Guidance and Public Disclosures to Implement Initial Margin Proposals, the May 2026 consultative document (d232) that codifies how central counterparties must disclose initial margin responsiveness, override frameworks, and simulation tools.

The RegLeg Brief Specialist Panel tested both models on the consultation's text and on the January 2025 BCBS-CPMI-IOSCO report it implements, and documents findings in which the models softened mandatory CCP obligations into discretionary "should consider" language, invented a three-category taxonomy the source report does not contain, and fabricated a three-element public disclosure structure for override frameworks the guidance does not enumerate.

Claude Opus 4.7, asked what obligation the final CPMI-IOSCO guidance places on CCPs to provide margin simulation tools, wrote that CCPs "should consider" making the tools available to clearing members and, where feasible, their clients. The 2024 consultative text from which the obligation derives reads: "Margin simulation tools with certain minimum functionality should be made available by CCPs to clearing members and their clients." The model converted a positive obligation into a discretionary consideration, the difference between a CCP being expected to provide a tool and a CCP being expected to think about providing one.

Asked about the structure of the January 2025 BCBS-CPMI-IOSCO report, Opus 4.7 also asserted the report's ten policy proposals fall into "three broad categories" of CCP transparency, governance and clearing-member transparency. The report's published text describes ten proposals aimed at resilience of the centrally cleared market ecosystem; no three-category taxonomy is stated in the source.

Claude Sonnet 4.6 reproduced the same obligation-softening on margin simulation tools and added a fabricated disclosure structure for the override framework. Asked what CCPs must publicly disclose about their override framework, Sonnet 4.6 enumerated three elements: instances or circumstances where overrides may be warranted, the key decision-makers authorised to exercise override discretion, and the permissible types of adjustments. The consultative text says only: "CCPs should publicly disclose relevant information on their override framework." The three-element list is the model's construction, not the regulator's.

A CCP risk officer, clearing-member compliance lead, or supervisor drafting a comment letter, board paper, or implementation plan against either output would understate CCP obligations on simulation tools, structure their override-framework disclosure against a taxonomy that does not exist in the source, and cite a category framework for the underlying policy proposals that the BIS press release does not endorse. That is the failure mode these findings document.