Source-Credit Fabrication on MAS Notice 637 (Amendment) 2025. Two frontier AI subjects tested by the RLB Specialist Panel produced confident, lawyer-shaped answers on the MAS Notice 637 risk-based capital adequacy for Reporting Banks that do not reconcile to the regulator-issued source. One invented an FHC-numbered MAS notice that does not exist; the other misread the regulator's own amendment-drafting convention. Both events sit in the same exposed-fabrication class.

The pattern in one line

For Risk functions at Singapore retail-banking divisions the pattern is clean: leading AI assistants will name a MAS instrument with the wrong number and will read a regulator-stated drafting convention against the cover note, in confident prose, with no flag of low confidence.

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 functions at singapore retail-banking divisions use AI for: regulatory-perimeter mapping for the MAS Notice 637 (Amendment) 2025 amendment cycle, and amendment-package reading where the cover note carries the controlling drafting convention. Each question is recorded against verbatim regulator-issued source text held by the Panel as primary substrate.

The Panel asked two frontier AI subjects to answer in their normal application register, then bound each answer to the corresponding paragraph of MAS Notice 637 and the corresponding cover-note language in the MAS Notice 637 amendment package. Where the AI assertion contradicted the verbatim regulator text the Panel recorded a FABRICATED_FACT finding under exposed_fabrication.

What the models got wrong

Finding 1: Fabricated "Notice FHC-N637" for financial holding companies. Opus 4.7 asserted that financial holding companies incorporated in Singapore are covered by a separate notice typically called "Notice FHC-N637 (Risk Based Capital Adequacy Requirements for Financial Holding Companies)". The MAS Notices and Directives register does not list any such instrument. The actual MAS Notice 637 applies, by paragraph 1.1, to Reporting Banks issued under the Banking Act; FHCs sit under a separate MAS notice issued under the Financial Holding Companies Act.

Claude Opus 4.7 produced the assertion in response to a question about how MAS Notice 637 applies to financial holding companies; Claude Sonnet 4.6 was tested as the comparator subject in the broader Panel run. The AI output reads as a lawyer-shaped citation, but the cited notice number is not on the MAS Notices and Directives register. Citation ID RLB-H-SG-MAS-NOTICE-637-CAPITAL-ADEQUACY-BANKS-2025-Q010-Opus47.

Finding 2: Misrepresented yellow-highlight meaning in MAS amendment PDFs. Opus 4.7 characterised the yellow highlighting in the MAS Notice 637 amendment PDF as a drafting or visual aid drawing attention to defined terms, cross-references, or items to read carefully. The MAS amendment package itself states on its face that yellow-highlighted text is annotation describing the change, and will not appear in the published untracked Notice. The AI characterisation contradicts the regulator's own cover note. Claude Opus 4.7 produced the yellow-highlight characterisation in response to a Specialist Panel application-style question on how to read the MAS Notice 637 amendment package.

The reading directly contradicts the cover-note convention on the face of the MAS PDF. Citation ID RLB-H-SG-MAS-NOTICE-637-CAPITAL-ADEQUACY-BANKS-2025-Q012-Opus47.

Why this matters for Risk

Risk at a retail-banking division translates the MAS Notice 637 framework into RWA computations, capital-buffer calibration, and stress-test scenarios for consumer-credit portfolios. Each artefact has to map cleanly to the regulator-issued source: the right instrument for the right entity, and the right reading of the amendment drafting conventions.

Both AI hallucinations tested by the RLB Specialist Panel attack that mapping. The fabricated "Notice FHC-N637" would, if relied on in a regulatory-framework map, drive RWA classification through an instrument that does not exist on the MAS Notices and Directives register. The amendment yellow-highlight misreading would lead risk to embed annotation text into templates and model documentation, producing a versioning artefact that the regulator will not reproduce in the published Notice.

For a risk function operating under MAS supervision the consequence is direct: a model-governance finding and a re-validation cycle that touches every downstream artefact.

The regulator's actual position

On the FHC perimeter, MAS Notice 637 states on the face of the Notice: "This Notice is issued pursuant to section 55(1) and section 65(2) of the Banking Act and applies to all Reporting Banks. [...] 11.2.2 A Reporting Bank need not comply with the requirements in this Part if it is a subsidiary of (a) another Reporting Bank which is subject to the requirements in this Part; or (b) a financial holding company which is subject to requirements similar to that set out in this Part." The Reporting Bank perimeter is therefore explicit in the Notice itself, and the FHC framework sits under a separate MAS notice issued under the Financial Holding Companies Act, not under a sibling 637-number instrument.

On the amendment yellow-highlight convention, the MAS Notice 637 amendment package states on its cover: "Text which is highlighted in yellow are annotations to describe changes, and will not appear in the published untracked version of MAS Notice 637. For instance, where amendments have been made to a selected paragraph of an Annex, only that paragraph will be reflected in this document, prefaced with the following explanatory text in yellow highlights: [Amendments to paragraph xx]." The convention is regulator-stated and controlling; any reading that treats the yellow text as substantive new Notice content is against the regulator's own cover-note language.

What this tells us about AI for Risk

The lens for retail-banking risk is exposed fabrication producing model-governance risk. The Specialist Panel categorised both findings as FABRICATED_FACT, which in a risk function translates into framework-map error and model-documentation error.

For the function the implication is that AI-drafted regulatory-framework content cannot enter the model-governance estate until each MAS instrument reference and each drafting-convention reading has been verified against the regulator-issued source. AI tools that do not enforce verbatim source binding will reproduce this class of error, and the propagation path through model documentation and stress-test scenarios is fast.

What the RLB Specialist Panel is doing about it

The Panel records each error against the verbatim regulator-issued source, names the AI subject, and binds the finding to a specific paragraph of MAS Notice 637 or the specific cover-note language in the amendment package. Findings are released publicly with the Citation IDs, the substrate document reference, and the operational consequence for each affected audience. Where firms want to take the Panel's verified primary substrate inside their own AI workflow, the Panel runs a partnership track that supplies the substrate package, the verification protocol, and a per-finding regression test set.

The partnership track is the operational answer to the class of failure on display in RLB-H-SG-MAS-NOTICE-637-CAPITAL-ADEQUACY-BANKS-2025-Q010-Opus47 and RLB-H-SG-MAS-NOTICE-637-CAPITAL-ADEQUACY-BANKS-2025-Q012-Opus47.

What Risk teams should do

Three actions follow for retail-banking risk.

First, every MAS instrument reference in an AI-drafted framework map should be resolved against the MAS Notices and Directives register before the artefact is released to model governance. The fabricated "Notice FHC-N637" would have failed this step; the FHC framework should be named as MAS's separate notice under the Financial Holding Companies Act.

Second, any AI-drafted reading of the amendment drafting convention should be checked against the MAS cover note, and model documentation should quote the cover note when it relies on the convention. AI gloss should be rejected.

Third, the risk function should hold a short AI-output verification protocol for regulatory-capital documentation: instrument name verified against MAS register, paragraph cited verified against published Notice, drafting-convention reading verified against cover note, version reconciled against the consolidated Notice. The RLB Specialist Panel can supply the protocol and the verified primary substrate.

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: MAS Notice 637 (Amendment) 2025, Risk Based Capital Adequacy Requirements for Banks Incorporated in Singapore · Substrate documents: mas-notice-637-amendment-2025.pdf, mas-notice-637-effective-2025-12-31.pdf · MAS portal: mas.gov.sg

Citation IDs referenced:

• RLB-H-SG-MAS-NOTICE-637-CAPITAL-ADEQUACY-BANKS-2025-Q010-Opus47
• RLB-H-SG-MAS-NOTICE-637-CAPITAL-ADEQUACY-BANKS-2025-Q012-Opus47