Risk functions at Singapore retail-banking divisions are increasingly using AI to update the retail-banking regulatory-capital framework map, draft RWA classification notes for the 31 December 2025 amendment, generate stress-test documentation for consumer portfolios under MAS Notice 637, and prepare board-level dashboards on FHC-level capital obligations. In Singapore-incorporated banks and financial holding companies the workflow shape is now consistent: a frontier AI assistant produces a clean first draft on MAS Notice 637 risk-based capital adequacy for Reporting Banks, and the reviewer is asked to spot-check the cited MAS instruments and drafting-convention claims against the regulator-issued source before the deliverable goes out.
The two AI failures recorded by the RLB Specialist Panel sit precisely at that spot-check boundary.
Two frontier AI models tested by the RLB Specialist Panel on MAS Notice 637 (Amendment) 2025 produced FABRICATED_FACT errors against the regulator-issued source held as primary substrate. The first invented a sibling "Notice FHC-N637" for financial holding companies that does not appear on the MAS Notices and Directives register; the actual FHC capital framework is a separate MAS notice issued under the Financial Holding Companies Act.
The second misread the yellow-highlight convention in the MAS Notice 637 amendment PDF as visual emphasis, when the regulator's cover note states the yellow is annotation describing the change and will not appear in the published untracked Notice. Both findings sit in the same failure class: Source-Credit Fabrication, where the AI produces a confident, lawyer-shaped citation that does not exist or contradicts a regulator-stated convention. Neither AI subject hedged, flagged low confidence, or refused.
Both produced clean, deployable prose with the wrong substantive content, which is the version of AI failure that is hardest for a reviewer to catch on a fast-moving deliverable. Questions are prepared by the RLB Specialist Panel based on real practical AI usage in the workflows the respective audience uses AI for. The Panel binds each AI finding to verbatim regulator-issued source text held as primary substrate, and records the AI subject, the question class, and the operational consequence for each affected audience.
For Risk functions at Singapore retail-banking divisions the operational consequence is concrete. A retail-banking regulatory-framework map that names a fabricated MAS notice would drive RWA classification and capital-buffer calibration through an instrument that does not exist. Risk-model documentation that captures amendment annotation as substantive Notice text would generate a versioning artefact that fails reconciliation against the published Notice. Both errors are direct model-governance issues.
The RLB Specialist Panel records each error against the underlying regulator-issued text and names the AI subject for audit transparency. The two findings carry Citation IDs 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; Claude Opus 4.7 is the AI subject in both events and the source-text excerpts are quoted verbatim in the briefing body that follows.
This is the consolidated view of findings. Click the Citation IDs or 'see details →' on any item for the full details for each finding.
Retail-banking risk teams responsible for the regulatory-capital view of a Reporting Bank that sits inside an FHC structure need precise framework references. Opus 4.7's fabrication of "Notice FHC-N637" would, if relied on in a regulatory-framework map, drive RWA classification and capital-buffer calibration through a non-existent instrument and produce limit and stress-test outputs that cannot be reconciled to any MAS source. Risk should source every notice reference from the MAS register; the FHC-level framework is MAS's separate notice under the Financial Holding Companies Act.
Retail-banking risk teams reading the amendment package to update capital and risk-monitoring templates need to recognise which text enters the consolidated Notice on commencement. Opus 4.7's reading of the yellow as visual emphasis would lead risk to embed annotation text into templates and model documentation, generating a versioning artefact that the regulator will not reproduce in the published Notice. The cover-note annotation convention controls; risk documentation should follow it directly.
Every finding on this page compares an AI subject's account of the rule against the regulator's verbatim text from the regulator's own portal. Both are linked. Each delta, its root causes, and impact analysis are documented and published with immutable Citation IDs.