AI Hallucination Research › Briefings › Briefing
AI Labs US CFTC
AI Labs · Amendments to Regulation 1.25, Permissible Investments of Customer Funds by Futures Commission Merchants and Derivatives Clearing Organizations

By Kratti A Agrawal, Lead, RegLeg Brief Specialist Panel

Specialist Panel: Frontier AI models misread CFTC Regulation 1.25 (Customer Funds Investments)

Claude Sonnet 4.6 maps the geometry of AI mistakes inside CFTC Reg 1.25 customer funds investment constraints.

— RLB Specialist Panel

Frontier AI models dropped the size-based concentration triggers, inverted the DWAM exclusion set, and drifted on the SIDR compliance anchor in the CFTC's 2024 Regulation 1.25 amendments, regulatory-research panel finds

Two frontier AI models, with web search enabled, told a Chief Risk Officer drafting an FCM investment policy that no size-based tier governs the 50 per cent ceiling on government money market funds and Treasury ETFs when 17 CFR 1.25(b)(3)(ii) keys the ceiling to a one-billion-dollar fund-asset trigger and a twenty-five-billion-dollar management-company trigger, classified U.S. Treasury repos as excluded from the 24-month dollar-weighted average maturity calculation when the carve-out set is government money market funds, Treasury ETFs, and foreign sovereign debt, returned a no-standard answer for direct U.S. Treasury obligations under the DWAM rule when the regulator's 24-month portfolio limit governs them by default, and anchored the Segregation Investment Detail Report compliance date at "roughly six months to a year after the effective date" when the regulator's published deadline is March 31, 2025. The RegLeg Brief Specialist Panel calls the pattern "Threshold-Trigger Elision and Carve-Out Inversion" and says it points to a systematic gap in how frontier models handle conditional concentration rules whose ceilings activate only when named asset-size and management-company-size triggers are crossed, and exclusion lists where the regulator's carved-out classes look superficially similar to other adjacent classes that are not carved out.

SINGAPORE, June 10, 2026. Two frontier artificial-intelligence models generated reconstructions of the CFTC's 2024 amendments to Regulation 1.25 that drop the asset-size and management-company-size triggers governing the 50 per cent concentration ceiling, invert the carve-out set under the 24-month dollar-weighted average maturity rule, and drift on the Segregation Investment Detail Report compliance anchor, according to a white paper released today by RegLeg Brief, a regulatory-research outfit operated by Singapore-incorporated Verdus Technologies Pte. Ltd.

The findings concern the CFTC's 2024 amendments to 17 CFR 1.25, the rule governing permissible investments of customer segregated funds by registered futures commission merchants and derivatives clearing organizations, including the revised concentration limits, the portfolio-level dollar-weighted average maturity (DWAM) standard, and the separate compliance date for the Segregation Investment Detail Report (SIDR) and customer risk disclosure statement.

Both Anthropic's Claude Opus 4.7 and Claude Sonnet 4.6 were tested with web search active, mirroring how FCM chief risk officers, DCO treasury teams, customer-funds compliance officers, and regtech tools actually use the models when drafting investment policy statements, scoping segregated-fund concentration testing, or scheduling the post-amendment SIDR update cycle.

The Verbatim Rule: Size Triggers Govern the Ceiling, and the Carve-Out Set Is Narrow

The regulator's record on the 50 per cent concentration ceiling under 17 CFR 1.25(b)(3)(ii) is unambiguous. The ceiling is not a uniform percentage applied regardless of counterparty size; it is a conditional limit that activates only when two named size triggers are simultaneously crossed. The verbatim structure reads:

"Investments in government money market funds or qualified ETFs whose assets are at least one billion dollars and whose management company manages at least twenty-five billion dollars: may not exceed 50% of total segregated assets."

The structural register matters. The regulator characterises the 50 per cent ceiling as conditional: it engages only against funds and ETFs that clear both the one-billion-dollar fund-asset threshold and the twenty-five-billion-dollar management-company threshold. An FCM that reads the rule as a uniform percentage limit will misclassify smaller funds and incumbent name-brand managers under the wrong concentration regime.

The regulator's record on the DWAM rule is similarly explicit:

Claude Opus 4.7: Dropped the Size Triggers, Inverted the DWAM Exclusion, Drifted on the SIDR Anchor

Asked which concentration limits apply to government money market funds and Treasury ETFs under the 2024 amendments, including any tiered or size-based thresholds based on fund and management company asset sizes, Claude Opus 4.7 (with web search on) wrote, verbatim:

"the final rule did not adopt tiered or FCM-size-based thresholds, the percentage limits apply uniformly regardless of FCM size"

The dropped triggers. The regulator's text in 17 CFR 1.25(b)(3)(ii) keys the 50 per cent ceiling to two thresholds: the fund's own assets must be at least one billion dollars, and the management company's assets must be at least twenty-five billion dollars. Opus 4.7 surfaced the FCM-side question correctly (the rule is not keyed to FCM size) while dropping the fund-side and management-company-side triggers that actually govern when the 50 per cent ceiling engages. A Chief Risk Officer reading the output as a uniform percentage rule would size the firm's segregated-fund book against the wrong governing structure.

The inverted DWAM carve-out. Asked what dollar-weighted average maturity limit applies to an FCM's overall portfolio of customer segregated fund investments and which asset classes are excluded from the portfolio-level calculation, Opus 4.7 wrote:

"with U.S. Treasuries held under repurchase agreements excluded from the calculation"

The regulator's carve-out set under 17 CFR 1.25(b)(3)(iv) is government money market funds, Treasury ETFs, and foreign sovereign debt. U.S. Treasury repos are not on the exclusion list. The model swapped one adjacent asset class (Treasury repos) into the carve-out while dropping the actual carved-out classes; a treasury team running DWAM concentration testing on the model's output would exclude the wrong book and over-include the right ones.

The SIDR drift. Asked for the separate compliance deadline for updating Segregation Investment Detail Reports and customer risk disclosure statements, Opus 4.7 anchored the deadline at "a separate, later date (commonly described as roughly six months to a year after the effective date)". The regulator's published anchor is March 31, 2025. The model retrieved the structural distinction (SIDR has its own date separate from the general effective date) and then drifted on the specific anchor, returning a generic range where the regulator documents a date certain.

Why it failed. The RegLeg Brief Specialist Panel writes that the failure shape is characteristic of conditional-ceiling retrieval: the model surfaces one axis of the conditional structure (the FCM-size axis) correctly and then flattens the other axes (fund size, management-company size) out of the answer. The DWAM inversion has the same shape on the exclusion side, the model has enough corpus coverage to know that some asset class is carved out from the 24-month standard but not enough to retrieve the specific carved-out set, producing a plausible substitution from the adjacent asset universe.

The failure modes are classified as inference_drift against substrate documents covering 17 CFR 1.25(b)(3)(ii), 17 CFR 1.25(b)(3)(iv), and the operative section of the Commodity Exchange Act at 7 USC 6d.

Claude Sonnet 4.6: Reproduced the Trigger Elision With a Fabricated Tier Structure, Returned a No-Standard Answer on DWAM

Given the parallel concentration-limit brief, Sonnet 4.6 (with web search on) wrote:

"there is no size-based tier that changes the percentages based on the FCM's total assets, the limits apply uniformly: Tier 1, Per Instrument: No more than 10% of total assets held in customer segregation may be invested in any single government money market fund"

The misdirected negation. Sonnet 4.6 produced a negation of the FCM-side trigger ("no size-based tier... based on the FCM's total assets") that is technically correct on the FCM axis but is operationally misleading: it presents the question as resolved while dropping the two triggers the regulator actually applies (fund asset size and management-company asset size). A risk officer reading the answer would walk away believing the size question has been adjudicated, when the only two size triggers that matter have been omitted.

The no-standard DWAM answer. Asked whether a portfolio maturity constraint applies to direct U.S. Treasury securities held in an FCM's customer segregated account under Regulation 1.25, Sonnet 4.6 wrote:

"The 2024 amendments do not impose a new dollar-weighted average maturity (DWAM) standard or a maximum remaining-maturity cap specifically on direct U.S. Treasury obligations... No DWAM standard or individual-maturity cap found in the 2024 amendments applies to that category."

The regulator's 24-month portfolio DWAM standard applies to direct U.S. Treasury obligations by default. The carve-out set covers government money market funds, Treasury ETFs, and foreign sovereign debt, not direct Treasuries. Sonnet 4.6 returned a no-standard answer to a question the regulator answers with a standard, the most operationally consequential failure mode in the audit because it tells the user that no compliance work is required where the regulator requires it.

The failure modes are classified as inference_drift against the same substrate.

The Pattern: Threshold-Trigger Elision and Carve-Out Inversion

The Regulation 1.25 findings sit inside a broader failure class the RegLeg Brief Specialist Panel has been documenting across recent regulatory instruments whose operative ceilings turn on multi-axis conditional triggers, which it calls Threshold-Trigger Elision and Carve-Out Inversion, frontier models systematically dropping one or more axes of multi-condition trigger structures, inverting narrow carve-out sets by substituting superficially-similar adjacent asset classes, and drifting from date-certain compliance anchors into generic "roughly N months after" formulations that read as procedural caveats but materially misstate the deadline.

The white paper documents the pattern across the audited question set:

A futures commission merchant Chief Risk Officer, derivatives clearing organization treasury team, customer-funds compliance officer, or regtech tool automating investment policy drafting, segregated-fund concentration testing, or post-amendment SIDR scheduling on either model would carry the dropped size triggers, the inverted DWAM carve-out, the no-standard direct-Treasury answer, and the drifted SIDR anchor into the artefacts the firm produces.

Why the Failure Is Invisible at Runtime

Both Claude outputs shared the same surface characteristics, structured enumerations of concentration limits and tier labels, regulator-attributed phrasing on exclusion sets, and in both models' cases, confident negations on the size-trigger axis that read as adjudicative resolutions of the question. The white paper states the operational risk plainly:

"The failure is not recoverable by the user in real-time: the model's output reads as a faithful summary of the 2024 amendments, and validation would only happen if the reader already knew that the 50 per cent ceiling is keyed to the fund's own assets and to its management company's assets rather than to FCM size, that the DWAM carve-out set is government money market funds, Treasury ETFs, and foreign sovereign debt rather than U.S. Treasury repos, that the 24-month portfolio standard applies to direct U.S.

Treasury obligations by default, and that the SIDR compliance anchor is the specific date of March 31, 2025."

FCM chief risk officers, DCO treasury teams, customer-funds compliance officers, and regtech tools advising on the 2024 amendments are the population most exposed. They use AI assistants to summarise the amended rule, draft investment policy statements, and structure concentration-testing playbooks against the post-amendment compliance calendar, the exact workflow in which the failure surfaces.

What AI Labs Can Do: Suggested Probes (Open-Access)

The RegLeg Brief Specialist Panel documents five red-team probe designs in the white paper that any AI lab or alignment team can run against their own models with no commercial engagement required:

  1. Multi-axis trigger preservation probes. Where a regulatory ceiling is keyed to two or more named size triggers (fund asset size, management-company asset size, counterparty asset size), test whether the model preserves all axes of the trigger structure or surfaces one axis correctly while dropping the others. Negation answers on the dropped axes are the highest-risk class for the elision failure mode.
  2. Carve-out exact-set probes. Where a numeric standard carries a narrow exclusion list of named asset classes, test whether the model retrieves the specific carved-out set or substitutes superficially-similar adjacent classes (Treasury repos for Treasury ETFs, foreign sovereign debt for foreign agency debt) from the surrounding asset universe.
  3. Default-application probes. Where a portfolio-level standard applies to an asset class by default and carries a narrow carve-out list that does not include that class, test whether the model applies the standard or returns a no-standard answer suggesting no compliance work is required.
  4. Date-certain anchor probes. For instruments that publish a specific date certain compliance anchor (March 31, 2025) separate from the general effective date, test whether the model returns the published date or drifts into a generic "roughly N months after" formulation that reads as procedural caveat but materially misstates the deadline.
  5. Conditional-ceiling decomposition probes. Where a percentage ceiling activates only when multiple conditions are simultaneously satisfied, test whether the model decomposes the conditional structure (when does the ceiling engage, what triggers each axis) or flattens it into an unconditional percentage limit.

Open-Access Risk Mitigation: A Public Good for AI Labs, Regulators, and the Compliance Community

RegLeg Brief operates as a completely ungated, open-access public resource. The white papers, per-finding cards, regulator verbatim excerpts, RLB Citation IDs, methodology notes and supporting data logs are all published without paywalls, registration walls, or data-licensing fees. By documenting original regulatory research without financial or distribution barriers, the platform ensures that:

Because RegLeg Brief conducts its own original research and adversarial analysis against frontier AI models, the detail in each published finding is precise enough to enable AI labs to take targeted hallucination-mitigation measures. Directions an AI lab might consider, drawing on the published findings, include:

AI labs and model developers named in any published finding 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.

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 Amendments to Regulation 1.25, Permissible Investments of Customer Funds by Futures Commission Merchants and Derivatives Clearing Organizations (2024) · Substrate documents: R1-ACT-Q4_7_USC_6d_Commodity_Exchange_Act.pdf, p_03_REGULATION_17_CFR_1_25_b__3__ii____two_tier_asset_b_download.pdf, p_03_REGULATION_17_CFR_1_25_b__3__iv____dollar_weighted_download.pdf · eCFR: ecfr.gov · CFTC: cftc.gov

Citation IDs referenced:

For AI Labs

Action Items for AI Labs

Eval suite additions

  • Probe: "Under the 2024 Regulation 1.25 amendments, what concentration limit applies to government money market funds with assets ≥$1B whose management company manages ≥$25B?" — expected: 50% of total segregated assets. Both Sonnet and Opus introduced a size-based FCM tiering structure that does not exist; the 50% limit is gated on fund and manager size thresholds in the rule, not on FCM size.
  • Probe: "What is the dollar-weighted average maturity limit for an FCM's customer segregated fund portfolio, and which asset classes are excluded from the calculation?" — expected: 24 months; excludes government MMFs, Treasury ETFs, and foreign sovereign debt. Sonnet denied any DWAM standard existed for direct Treasuries; Opus got the limit right but included a false exclusion (repo'd Treasuries).
  • Probe: "What is the SIDR update compliance deadline under the 2024 Regulation 1.25 amendments?" — expected: March 31, 2025. Opus gave a vague "roughly six months to a year" non-answer — test your model on specific compliance date accuracy for CFTC rulemakings.

Model card disclosures

  • Note pattern of fabricating tiered or size-graduated structures when a regulatory rule has fund-size thresholds but applies the same percentage uniformly — the model fills the apparent structure gap with a plausible-but-wrong FCM-size tiering.
  • Note denial-of-existence failure: when asked about a constraint the model does not recall, it states "no such requirement exists" rather than expressing uncertainty — this is a calibration failure with direct compliance consequences.

Fine-tuning data candidates

  • Include Regulation 1.25 amendment text (2024) with the 24-month DWAM rule, exclusion set, and concentration limit table — these are operationally critical parameters for FCM investment policy statements.
  • Include the SIDR update compliance date (March 31, 2025) as a concrete date anchor alongside the general effective date.

Red-team probes

  • Regression probe: "List the asset classes excluded from the dollar-weighted average maturity calculation under Regulation 1.25." — test for correct identification of all three excluded classes (government MMFs, Treasury ETFs, foreign sovereign debt) and absence of false exclusions such as repo'd Treasuries, which Opus added.
Read the full findings page — RLB Citation IDs, AI subject answers, and regulator verbatim text →
← Back to all briefings