Executive Summary
This cell examines how AI tools perform when Risk teams at US hedge funds query the CFTC's December 2025 digital asset collateral framework — specifically the Market Participants Division's no-action relief and tokenized asset staff guidance that reshapes permissible collateral composition and haircut methodology for FCMs and their customers. Across the questions tested, AI tools produced at least one materially wrong answer on this regulation. The failure identified involves AI presenting the 20% minimum haircut floor as the operative rule for all customer-posted digital asset margin calculations — ignoring the governing multi-DCO rule that displaces it.
For a Risk function navigating collateral acceptance policies, margin sufficiency controls, and customer disclosure frameworks under this guidance, that omission is not a gap in nuance: it produces a structurally incorrect haircut methodology that would pass internal review undetected.
How AI gets this regulation wrong
The AI failure we found on this regulation involves inventing a rule — presenting a floor requirement as though it were the complete governing standard and suppressing the operative hierarchy that actually controls the calculation. The effect is an answer that is technically accurate in isolation but structurally wrong in application, because it omits the condition under which the floor is overridden.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Misstated Rule | 1 | Finding#1 |
What that means for your team
The practical risk for Risk functions at US hedge funds concentrates in wrong deliverables — policies, haircut schedules, or internal guidance documents built on a margin calculation framework that doesn't reflect what the CFTC actually requires. For a team whose collateral controls feed directly into margin sufficiency attestations, FCM counterparty oversight, and customer disclosure obligations, an incorrectly constructed haircut methodology is the kind of error that survives multiple review rounds before surfacing under regulatory scrutiny.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects your department
Risk teams at US hedge funds engage this guidance at the FCM interface — when reviewing the collateral acceptance terms their prime brokers and FCM counterparties operate under, when assessing whether digital assets posted as initial margin meet the framework's requirements, and when building or reviewing internal policies that govern what collateral the fund itself can pledge.
The guidance is also directly relevant when Risk is scoping a new crypto-native product or fund structure where digital asset collateral is a core operational feature, and when a business line asks Risk to confirm that a proposed margin arrangement is consistent with the CFTC's current posture.
The specific failure risk in this guidance is concentrated in the haircut methodology. When a hedge fund's Risk team asks AI to clarify the margin haircut rules for digital assets accepted by multiple registered clearing organizations at different rates, the AI's answer determines how the team constructs the margin sufficiency threshold used in counterparty exposure models and internal controls.
An answer that only describes the 20% floor — without surfacing the governing rule that the highest DCO haircut applies when multiple DCOs accept the same asset — produces a haircut schedule that systematically underestimates margin requirements wherever any DCO prices the asset above 20%.
If that methodology reaches the firm's collateral control framework, it introduces a structural undershoot into the fund's margin adequacy calculations. At position scale, the gap between applying the floor versus applying the highest DCO haircut can represent meaningful under-margining — an exposure that neither the fund's internal controls nor standard FCM netting would catch until a margin call reveals the discrepancy. The regulatory consequence is not abstract: under CFTC enforcement, margin calculation errors in customer collateral contexts carry direct liability for the FCM and reputational and legal exposure for the hedge fund operating on incorrect internal guidance.
The findings at a glance
One finding was identified across AI testing on this regulation for Risk teams at US hedge funds. It targets the specific multi-DCO haircut hierarchy that governs customer-posted digital asset margin calculations.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | Multi-DCO haircut hierarchy displaced by floor-only rule | Hallucination | RLB-F-US-CFTC-DIGITAL-ASSET-COLLATERAL-TOKENIZED-ASSETS-STAFF-GUIDANCE-2025-Q007 |
Aggregate impact
The single finding on this regulation is narrow in scope but high in operational consequence. It sits at the intersection of two conceptually adjacent requirements — the FCM's customer margin haircut rule and the firm's proprietary capital charge — and AI tools tested confused the floor case with the governing case. The 20% minimum haircut for digital assets not accepted by any DCO is accurate; the error is treating that floor as though it governs all margin calculations, including situations where one or more DCOs do accept the asset but at divergent rates.
For Risk teams, the failure pattern is particularly hazardous because the AI answer is not obviously wrong. It cites a real number from the real guidance, produces a plausible methodology, and presents no internal contradictions. A reviewer who is not already familiar with the multi-DCO hierarchy — precisely the junior analyst or business-side recipient for whom AI output carries the most risk — has no reason to question the answer. The error only becomes visible if the reviewer independently cross-references the specific multi-DCO provision, which is exactly the step the AI tool rendered unnecessary in the first place.
The systemic exposure for a US hedge fund is that this failure is structurally replicable. Any collateral policy, margin model documentation, or FCM counterparty assessment that was drafted using AI on this regulation faces the same omission. Where firms have multiple digital asset positions across FCMs that each accept the same assets at different haircut rates, the margin adequacy picture derived from an AI-guided 20% floor methodology is wrong for every one of those positions simultaneously.
Risk functions that have not independently verified the multi-DCO rule should treat any existing AI-assisted work product on this topic as requiring technical re-check before it supports a regulatory attestation or counterparty control sign-off.
What your team should do
The default position for Risk teams using AI on this guidance should be: verify the haircut methodology independently before any work product leaves the Risk function. The December 2025 guidance is recent enough that AI training coverage is uneven, and the multi-DCO haircut hierarchy is precisely the kind of conditional rule that AI tools flatten into its simplest version.
Any AI-generated description of the margin haircut framework should be checked line-by-line against the CFTC Market Participants Division's published text, not against secondary commentary or press coverage — the finding in this cell traces directly to AI tools that cited press articles rather than the primary guidance, and the press articles did not reproduce the multi-DCO rule.
The specific control to add is a step in the policy and collateral-framework review process that explicitly confirms whether the asset in question is accepted by any registered DCO, and if so, what the highest applicable haircut rate is across all accepting DCOs. That determination needs to sit upstream of the 20% floor check in any margin calculation methodology — AI output that does not reflect this ordering is incomplete regardless of what else it gets right.
Where the fund's FCM counterparties have posted their own collateral schedules under the guidance, those schedules are the authoritative reference for the highest-DCO-haircut determination, not AI synthesis.
AI tools are reasonably safe on this regulation for orientation-level questions: understanding the structure of the guidance, identifying the two distinct 20% requirements and what they apply to at a conceptual level, or mapping which parts of the framework are relevant to a specific product or position type. The failure risk is specific to calculation methodology — wherever the output would feed directly into a margin number, a haircut schedule, or a control threshold, treat AI as a starting draft that requires primary-source verification, not a finished answer.
How RLB Can Help
RegLeg's published Hallucination Research gives your Risk team a concrete pre-flight check before you rely on AI-generated output for any regulatory question — margin calculations, reporting thresholds, capital treatment determinations. The findings catalogue where AI assistants confidently produce wrong numbers, mischaracterise regulatory scope, or invert the direction of a requirement. Running that catalogue against the specific regs your desk actually touches takes twenty minutes and tells you which outputs to verify closely and which workflows carry acceptable exposure.
That is not a compliance formality; it is operational risk management for a function that is already integrating these tools into the daily workflow.
On a bespoke basis, we map your firm's AI-supported Risk workflows against the failure modes we have catalogued across the regulators you report to — SEC, CFTC, NFA, and the relevant prudential perimeter. The output is a prioritised exposure inventory: which question types your analysts are currently asking AI tools, which of those question types have documented hallucination patterns, and where the gap between confident AI output and regulatory ground truth is wide enough to require a human check or a policy constraint.
Hedge fund Risk functions running AI-assisted VaR attribution, counterparty credit monitoring, or 4.22(a)/AIFMD look-through analysis each carry a distinct hallucination-risk profile; the mapping is fund-structure-specific, not generic.
We also offer a confidential review of your firm's existing AI-use policy against the failure-mode catalogue — typically identifying the categories of regulatory question that are either unaddressed or addressed with controls calibrated to the wrong risk level. Where gaps exist, we produce a prioritised remediation list your COO or General Counsel can action. Separately, we can develop training material and CPD-aligned content for your Risk team: structured enough to satisfy internal sign-off requirements, specific enough to be useful to people who already know what they are doing.