Numeric Drift and False-Negative Availability Claim on the CPMI API harmonisation recommendations for cross-border payments. Two frontier AI models tested by the RLB Specialist Panel produced regulator-fluent answers on CPMI's October 2024 d224 report that the panel was able to bind, line-by-line, to regulator-issued source text the AI subjects did not match. The pattern recurs across 2 substrate-bound findings on named partnerships, per-recommendation stakeholder taxonomy, ISO 20022 update language and fast payment system statistics.

The pattern in one line

For risk teams at corporate banking firms, the audit pattern is consistent: the AI subjects answer CPMI questions in the same fluent, structured register they use for verified facts, even when the underlying CPMI source the answer cites was never actually retrieved. The wrong number, the wrong availability claim, the wrong named partnership, and the fabricated date-and-format mandate all arrive in the same prose voice as the AI's correctly-sourced output.

There is no surface signal in the answer that tells the reader which paragraphs are bound to a CPMI publication and which are constructed from category names and adjacent training data. The failure pattern is therefore not primarily a prompt-engineering issue; it is a verification-process issue, and it sits upstream of any tone or formatting fix that risk teams at corporate banking firms apply at the AI drafting stage.

How the RLB Specialist Panel tested this

Questions are prepared by the RLB Specialist Panel based on real practical AI usage in the workflows risk teams at corporate banking firms use AI for. Each question is bound to verbatim regulator-issued source text held as primary substrate before the question is ever issued to a model. The Panel issued each question to two frontier AI models with web search enabled, recorded the answer verbatim with timestamp and model identity, and then bound the answer against the regulator-stated text from CPMI's primary publications.

The primary substrate for this regulation includes the October 2024 d224 report (the master CPMI API harmonisation recommendations and toolkit), CPMI Brief No. 9 (November 2025) on the pre-validation API recommendation and the SARB collaboration, and CPMI speech sp231115 (Tara Rice, November 2023) on the global fast payment system landscape statistics. Findings are only published where the AI answer fails the substrate-binding check against one or more of these primary CPMI publications. Positive observations (correctly-sourced answers, well-calibrated refusals) are not published as findings; only substantive regulator-contradicting errors enter the public audit record.

What the models got wrong

Claude Opus 4.7: Global FPS count understated by survey-sample substitution

Claude Opus 4.7 with web search stated that the 2025 monitoring survey covers 57 FPS as the most recent CPMI count of operational systems globally. The answer is recorded verbatim and cited as RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q010-Opus47, bound to the regulator's actual text:

CPMI speech sp231115 (Tara Rice, November 2023) records '70+ domestic fast payment systems currently operational globally', alongside '14 fast payment systems already enabling cross-border exchanges' and '24 systems planning linkages within five years'.

The substantive consequence: A market briefing or strategy memo that quotes 57 understates global connectivity by roughly 20 percent, mis-prioritises corridor strategy, and embeds an authoritative-sounding statistic that the primary source contradicts.

Claude Sonnet 4.6: FPS operational-split data falsely declared unavailable

Claude Sonnet 4.6 with web search replied that no precise percentage breakdown of central bank vs privately operated FPS is enumerated in the public Brief 10 summaries available. The answer is recorded verbatim and cited as RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q010-Sonnet46, bound to the regulator's actual text:

CPMI speech sp231115 (Tara Rice, November 2023) records: '40% operated by central banks; 35% by private entities', alongside the '70+ domestic fast payment systems currently operational globally' figure.

The substantive consequence: A briefing that accepts the false-negative claim either leaves a published CPMI data point off the deliverable or burns time chasing a primary-source gap that does not exist.

Why this matters for risk teams at corporate banking firms

A board-risk paper that records a CPMI cutover date the regulator never set is a factual error in a board-approved risk-appetite document. A risk dashboard that uses 57 rather than 70+ as the FPS connectivity baseline mis-sizes corridor exposure. An enterprise risk register entry recording 'no SARB pre-validation workstream identified' carries a verifiable error into a supervisory deliverable. The next supervisory testing on AI use in risk reporting will find these gaps. The factual gap, once it enters a deliverable that risk teams at corporate banking firms produce, is durable.

It travels into firm-level institutional records, into client-facing collateral, into supervisory horizon scans and into engagement-letter knowledge bases. It is hard to walk back without quietly issuing a correction, and the supervisory cost of the correction is often higher than the original verification effort would have been.

The regulator's actual position

The regulator-stated positions, drawn verbatim from CPMI's primary publications (d224 October 2024, CPMI Brief No. 9 November 2025, speech sp231115 November 2023), read as follows.

CPMI speech sp231115 (Tara Rice, November 2023) records '70+ domestic fast payment systems currently operational globally', alongside '14 fast payment systems already enabling cross-border exchanges' and '24 systems planning linkages within five years'.

CPMI speech sp231115 (Tara Rice, November 2023) records: '40% operated by central banks; 35% by private entities', alongside the '70+ domestic fast payment systems currently operational globally' figure.

Each of these passages is held inside the RLB primary substrate against which the AI answer was bound. Where the AI answer states a partnership, a stakeholder assignment, a dated commitment or a numeric figure that contradicts the verbatim regulator text above, the discrepancy is recorded as a substantive finding with an immutable RLB Citation ID.

What this tells us about AI for risk teams at corporate banking firms

Risk functions should read the audit as a control finding against AI-drafted regulatory inputs. The AI manufactures numbers and partnerships in the same prose register it uses for verified facts. A risk function that builds an explicit primary-source verification step into AI-drafted risk reporting controls the exposure. A function that treats AI-drafted CPMI figures as inputs to roll up into risk dashboards introduces factual error into the risk record at the speed of AI drafting.

The taxonomy lens for this audience is straightforward: treat the AI's regulator-quoting prose as a draft to be verified, not as a sourced deliverable to be word-smithed. The verification step has to be explicit, separate from the AI drafting step, and run against a named CPMI publication rather than against general web search.

The four substantive CPMI failure modes documented in this audit (named partnership denial, per-recommendation stakeholder fabrication, ISO 20022 structured-address mandate fabrication, FPS count and operator-split misstatement) are now publicly logged, so the verification step does not need to be open-ended: it can run against the known failure list as a starting point.

What the RLB Specialist Panel is doing about it

The RLB Specialist Panel records each CPMI finding with an immutable RLB Citation ID and binds it to the regulator-issued primary source. The CPMI API Harmonisation for Cross-Border Payments hub at https://reglegbrief.com/regulators/j1/int/bis-cpmi/cpmi-api-harmonisation-cross-border-2024/ is the live record of substrate-bound CPMI AI failure modes that risk teams at corporate banking firms can lift directly into firm-level AI-use risk registers, into client-deliverable verification checklists, and into supervisory horizon scans. The Panel partners with firms, agencies and standards bodies that use AI on CPMI deliverables, surfacing new failure modes against the firm's own actual deliverable types as the CPMI issues further briefs and speeches.

Partnership engagement gives a firm or agency early access to flagged AI failure modes against its own deliverable templates, and a documented audit trail showing that the firm has tested for the failure pattern before a supervisor asks about it.

What risk teams at corporate banking firms teams should do

The action items below are framed for risk teams at corporate banking firms specifically and map onto the four documented failure modes:

• Add a primary-source check on every CPMI-sourced figure in a risk dashboard, board paper or enterprise risk-register entry, before the artefact is approved.
• Maintain a verified CPMI fact pack inside the risk function and reconcile any AI-drafted CPMI text against the fact pack as a default step.
• Update the firm's AI-use risk register to record the four documented CPMI failure modes as known exposures with named control owners.
• Brief risk leads on the SARB pre-validation partnership so that the named workstream is in the risk-function knowledge base.
• Stamp AI-drafted risk papers and dashboard updates with a sourcing tag identifying primary-source-verified figures against AI-asserted figures.
• Engage with the RLB Specialist Panel partnership track to surface CPMI AI failure modes against the risk function's deliverable types.

Each action is operationally lightweight at the unit-deliverable level and compounds across the function as a control standard. The cost of building the verification step into the AI drafting workflow is consistently lower than the cost of issuing a correction on an AI-drafted deliverable that has already entered firm or supervisory records.

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 Report d224, Harmonised application programming interfaces for enhancing cross-border payments (October 2024) · Substrate documents: p_10_GUIDELINE_Tara_Rice_speech_Nov_2023___FPS_statisti_d230.htm · CPMI portal: bis.org/cpmi

Citation IDs referenced:

• RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q010-Opus47
• RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q010-Sonnet46