AI Hallucination Research

RegLegBrief is the public research output of Verdus Technologies Pte. Ltd., a Singapore-incorporated regulatory technology firm. Through our consulting practice we see, every day, how Licensed Professionals (lawyers, compliance officers, risk officers, internal auditors, financial advisers) and department heads across sectors and sub-sectors use frontier AI on regulatory work. We see the same patterns of failure surface across regulations, across jurisdictions, across model families and Agentic systems.

We document those failures here, free and open-access, with a verbatim regulator excerpt for every claim and an immutable Citation ID for every finding. AI labs and model developers can use these findings to harden their models and retrieval architectures. Licensed professionals and department heads can use them to brief their own IT teams on where checks and controls belong in the AI tools and Agentic systems already in use across their floors.

A pattern emerged. After enough confirmed findings across enough rules, regulations, jurisdictions, and audience domains, the hallucinations are not random. They map to specific substrate conditions, specific prompt shapes, and specific retrieval gaps. A previously probabilistic outcome is becoming a predictable one, and a predictable failure mode is a designable safeguard. That is the real impact of the RLB Specialist Panel's work.

The catalogue spans capital adequacy, conduct and consumer protection, derivatives and swap dealers, market infrastructure (CCPs, CSDs, payment systems), cross-border payments and ISO 20022, sovereign-debt surveillance, and maritime treaties, verified against primary sources from BIS/CPMI, IMF, OECD, UN Treaty Collection, FCA, MAS, and CFTC. New regulations are added on a continuous cadence as they are sponsored or commissioned.