A frontier AI model substituted the base 20 per cent haircut for the multi-DCO worst-case selection rule on customer margin collateral.

A frontier AI subject tested by the RLB Specialist Panel described the haircut floor that applies when no registered DCO accepts the asset, and presented it as the operative rule for the multi-DCO scenario, where the highest accepted haircut actually governs.

The pattern in one line

A frontier AI model tested on the CFTC Digital Asset Collateral Framework described the base 20 per cent haircut floor as the operative rule for the multi-DCO scenario, when the regulator's selection rule is the highest accepted haircut among all registered DCOs that accept the asset, producing a risk answer that would systematically under-collateralise customer accounts.

How the RLB Specialist Panel tested this

The questions in this cell were prepared by the RLB Specialist Panel based on real, practical AI usage in the workflows that risk teams at hedge funds firms actually use AI for under the CFTC Digital Asset Collateral Framework. Each question targets a specific deliverable type where an AI assistant is plausibly the first draft: a memo, an eligibility paragraph, an onboarding checklist line, a haircut-model assumption, a regulator-facing filing sentence. The Panel issued each question to two frontier AI subjects with web search active.

The Panel then bound every AI response to verbatim regulator-issued source text held as primary substrate, comparing the model output against the CFTC staff letter text and the regulator-issued source documentation for each provision. Only responses where the AI subject was demonstrably wrong against the verbatim regulator-issued source text are published as findings; responses that were substantively correct, or that refused on calibration grounds, are retained internally and not surfaced.

What the models got wrong

Finding: Multi-DCO haircut tiebreaker rule replaced by the base 20 per cent floor. The Specialist Panel asked, in application form, how the customer margin haircut for digital assets is calculated when multiple registered DCOs each accept the same digital asset but at different haircut rates, a scenario that arises whenever an FCM clears the same crypto asset through more than one DCO, which is the commercially dominant pattern for bitcoin, ether, and the eligible payment stablecoins.

Claude Sonnet 4.6 with web search active answered that for digital assets not accepted by any registered DCO as initial margin, the FCM must apply a minimum 20 per cent haircut to the current market value of the customer-deposited collateral, and presented this base-case framing as the operative rule for the multi-DCO scenario the Panel asked about (RLB-H-US-CFTC-DIGITAL-ASSET-COLLATERAL-TOKENIZED-ASSETS-STAFF-GUIDANCE-2025-Q007-Sonnet46). The model's substantive description of the 20 per cent floor is correct as far as it goes: the staff letter does set a minimum 20 per cent haircut for digital assets that no registered DCO accepts as initial margin.

The model's error is in transplanting that base-case rule onto the multi-DCO conditional. The substrate held by the Panel records the regulator's actual selection rule: where multiple DCOs accept the same asset at different haircuts, the FCM must apply the highest such haircut. The 20 per cent floor the model described is the regulator's base case for the no-DCO-accepts-it scenario; it is not the operative rule for the multi-DCO scenario, where the worst-case selection rule governs.

A customer margin programme that operationalises the model's framing systematically under-collateralises customer accounts that hold the same digital asset across two or more DCOs at differing haircut rates, which is the realistic operating environment rather than the edge case.

Why this matters for Risk teams at Hedge Funds firms

For risk teams at hedge funds firms accepting or posting digital asset margin collateral under the CFTC Digital Asset Collateral Framework, the accuracy of the haircut methodology and the post-onboarding reporting map drives the firm's customer-collateral coverage and its ongoing regulatory standing. A haircut model built on the base 20 per cent floor instead of the multi-DCO highest-accepted-rate rule produces systematically light collateralisation across customer accounts that hold the same digital asset across multiple DCOs, an exposure that surfaces only in stress, by which point the under-collateralisation has been priced into the customer book for months.

A post-phase obligation map that drops the weekly digital asset reporting cadence creates a recurring reporting violation that accrues silently between regulator engagements. The risk team owns the model assumptions and the obligation map, and each of these errors translates directly into mis-priced customer-collateral exposure, regulatory enforcement risk, and an inaccurate stress-testing baseline. The downstream cost of correcting a haircut model assumption after the customer book has been onboarded under the wrong rule is materially higher than the cost of verifying the conditional selection rule against the staff letter before the model goes into production.

The regulator's actual position

Multi-DCO scenario: highest accepted haircut governs. The CFTC staff letter contains two distinct 20 per cent figures and they govern different scenarios. One is a haircut floor for digital assets that no registered DCO accepts as initial margin: in that case, a minimum 20 per cent haircut to current market value applies to customer-deposited collateral. The other 20 per cent applies to a separate proprietary-holdings concentration limit, a distinct rule on the firm's own book rather than the customer book. Neither of these is the operative rule for the multi-DCO conditional.

For the multi-DCO scenario where two or more registered DCOs each accept the same digital asset at different haircut rates, the operative selection rule is the highest such haircut, applied to the customer's posted collateral. The structural logic is the regulator's worst-case selection: where a market-recognised haircut range exists, the FCM applies the top of the range, not an arbitrary floor and not the lowest-DCO rate. The base 20 per cent floor is the rule for the no-DCO-acceptance case; it is not the rule for the multi-DCO case.

What this tells us about AI for Risk teams at Hedge Funds firms

For risk teams at hedge funds firms working with AI on the CFTC Digital Asset Collateral Framework, the recurring pattern is a Dropped-Qualifier Misstated Rule. The AI assistant described the base 20 per cent haircut floor, which is real, but presented it as the operative rule for the multi-DCO scenario, where the highest accepted haircut actually governs.

This is a class of error that routine citation-checking does not catch: the staff letter reference the AI cites is the right operative letter, the 20 per cent figure the AI cites does exist in the letter, and the substantive paraphrase reads as authoritative. The defect is conditional: the AI has applied a base-case rule to a conditional scenario where the regulator's worst-case selection rule governs. The defensive workflow that catches this is a conditional read of the full haircut rubric in the staff letter, including the multi-DCO selection rule, against the verbatim staff letter text.

The practitioner takeaway is operational: when an AI assistant offers a haircut answer for customer-posted digital asset collateral, always re-verify the conditional selection rule against the regulator-issued text before the output enters a customer-collateral model, an FCM customer-onboarding memo, or a customer-facing eligibility summary. The risk is not a citation slip; it is a systematic under-collateralisation of customer accounts in the realistic multi-DCO operating environment.

What the RLB Specialist Panel is doing about it

The RLB Specialist Panel is engaging with the AI subjects' developers and with practitioner audiences working under the CFTC Digital Asset Collateral Framework. The Panel maintains an audit register of confirmed hallucinations bound to verbatim regulator-issued source text, surfaces them on the live regulation page and on each audience-specific briefing, and accepts right-of-reply submissions from the AI subjects' developers and from regulator-side reviewers.

For risk teams at hedge funds firms this means the same questions can be re-issued against successor model releases; the bound substrate makes it straightforward to verify whether a specific failure mode has been corrected upstream, or whether the same hallucination is still being produced. Partnership briefings with AI labs are offered against the audit register, not against synthesised demonstrations, so the corrections that matter are evidenced against the operative staff letter text rather than against a paraphrase chain.

What Risk teams at Hedge Funds firms teams should do

For risk teams at hedge funds firms drawing on AI in workflows that touch the CFTC Digital Asset Collateral Framework, the practical action items are direct:

• Anchor every haircut assumption produced by an AI assistant against the operative CFTC staff letter haircut rubric, before the assumption enters a customer-collateral model or a risk committee paper.
• Apply the multi-DCO highest-accepted-rate selection rule when modelling customer margin haircuts on digital asset collateral; the base 20 per cent floor is the no-DCO-acceptance case, not the multi-DCO case.
• Treat the post-onboarding obligation map as a high-risk paragraph: read the operative staff letter directly and confirm the continuing weekly digital asset reporting cadence before incorporating the obligation map into a risk runbook.
• Maintain a documented log of AI-assisted CFTC staff-letter citations entering risk work product, with the verification step explicitly recorded.
• Track the RLB Specialist Panel audit register on the CFTC Digital Asset Collateral Framework so that the risk team is notified when a hallucination is added, withdrawn, or corrected against a successor model release.

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: CFTC Staff Advisory on Digital Asset Collateral and Tokenized Assets (2025) · Substrate documents: p_02_GUIDELINE_CFTC_Staff_Letter_25_40___26_05___two_di_download.pdf · CFTC: cftc.gov

Citation IDs referenced:

• RLB-H-US-CFTC-DIGITAL-ASSET-COLLATERAL-TOKENIZED-ASSETS-STAFF-GUIDANCE-2025-Q007-Sonnet46