The RLB Specialist Panel

Independent verification body for the RegLegBrief AI Hallucination Research catalogue.

TL;DR. The RLB Specialist Panel verifies every published finding against the regulator's verbatim text on the regulator's own portal. The Panel is independent of the AI subjects under test and independent of the regulators whose rules are tested. Verification operates under five codified rules, classification uses a seven-failure-mode taxonomy, and publication is gated by the no-substrate-no-audit rule. Every finding carries an immutable Citation ID and stays permanently on the record.

What the Panel is

The RLB Specialist Panel is the editorial and verification body that sits behind every published RegLegBrief finding. Its mandate has three parts:

Test named AI subjects (frontier model + tool variant — for example, Claude Opus 4.7 with web search, GPT-5 with web search) against regulator-issued primary sources on the regulator's own portal.
Classify every disagreement between the AI subject's account and the primary source by failure mode — deontic-register inversion, negation-reversal, schema substitution, entity misidentification, and other documented categories from the seven-mode taxonomy.
Publish the audit trail with an immutable Citation ID, the verbatim AI output, the verbatim primary-source text, the classified failure mode, and the URL to the regulator's portal.

Composition and experience

The Panel comprises senior regulatory specialists drawing on combined decades of practitioner experience across financial-services regulation, administrative law, and AI verification. Individual names are not published; the Panel speaks as a body. This is a deliberate editorial choice: every finding is verified against a primary source that any reader can re-verify independently, so the authority of a finding rests on the cited primary source and the documented methodology, not on the individual reviewer's name. The institutional accountability sits with Verdus Technologies Pte. Ltd., the operating publisher (Singapore, UEN 201616982R).

Domain coverage

The catalogue covers regulations across the following domains, with new domains added as new regulations are sponsored or commissioned:

Domain	Examples in the current catalogue
Banking capital adequacy & prudential	MAS Notice 637, CFTC FCM/DCO customer-funds investment rules
Conduct and consumer protection	FCA Consumer Duty (PS22/9 + PRIN 2A)
Derivatives & swap-dealer regulation	CFTC Swap-Dealer Business Conduct Documentation, CFTC Digital Asset Collateral / Tokenized Assets Staff Guidance
Market infrastructure (CCPs / CSDs / payment systems)	CPMI-IOSCO Principles for Financial Market Infrastructures (PFMI), CPMI-IOSCO PFMI Level 3 General Business Risk, CPMI-IOSCO Cyber Resilience FMI
Cross-border payments & messaging	CPMI ISO 20022 Harmonisation, CPMI API Harmonisation Cross-Border Payments
Margining & collateral	CPMI-IOSCO Initial Margin Disclosure Consultation
Sovereign debt & IMF surveillance	IMF Guidance on Financing Assurances in Sovereign Arrears, IMF Review of Charges & Surcharge Policy Reform
Maritime / environmental treaties	BBNJ High Seas Biodiversity Agreement

Browse the full regulator and regulation index →

Verification protocols (five codified rules)

The Panel operates under five codified editorial rules. These are enforced in the publication pipeline itself, not as best-practice guidance — the pipeline refuses to publish a finding that fails any of them.

Rule	What it means in practice
No-substrate-no-audit	The Panel does not generate questions, AI runs, or findings against documents whose verbatim text the pipeline cannot extract from the regulator's own portal. Wayback Machine fallback is permitted; secondary sources are not.
Publish-only-negative-findings	Published findings are negative only — the AI subject demonstrably diverged from the primary source. Positive observations (correct identification, well-calibrated refusal) are filtered at the publish gate. The catalogue is the record of where AI got it wrong, not where AI got it right.
Hallucinations-only-public	Blind spots (retrieval or tooling failures) are retained in the internal database but not published on public surfaces. Hallucinations (direct model errors against the primary source) are the only items published, because they are the items AI labs cannot argue down.
Major-conflicts-over-minor-wording	The Panel filters for material substantive conflicts — wrong entity, wrong number, wrong date, wrong scope, inverted position — not synonym substitutions or cosmetic title differences. Five sharp redlines beat twenty diluted ones.
Substrate-source attribution	Every published finding cites the substrate source rollup at the regulation level and per-redline. The reader can trace each delta back to the exact regulator-portal URL it was verified against.

Classification taxonomy

Failures are classified into a seven-mode hallucination taxonomy (covering how the AI's account diverges from the primary source) and a two-mode blind-spot taxonomy (covering why the AI could not access the source). The hallucination categories drive what is published; the blind-spot categories drive what stays internal. Modes include:

Deontic-register inversion — a "must" rendered as a "may," or vice versa.
Negation-reversal — a prohibition rendered as a permission, or vice versa.
Schema substitution — a wrong-shape answer (numeric where qualitative, jurisdictional where temporal).
Entity misidentification — the wrong regulator, body, document, or counterparty named.
Numeric drift — a threshold, ratio, or date misstated.
Scope inflation / contraction — the rule's applicability scope misstated.
Synthetic citation — a non-existent rule, paragraph, or version invented.

The four-way citation taxonomy (Pretextual / Inaccessible / Fabricated / Accurate-with-Context) categorises how each AI subject cited its source. See /methodology/ for the full taxonomy and the five-step research pipeline.

Independence

The Panel is independent in two specific senses:

Independent of the AI subjects under test. No AI lab compensates the Panel for findings against its model. The Panel publishes findings against frontier models from Anthropic, OpenAI, Google, and others on the same editorial terms.
Independent of the regulators whose rules are tested. No regulator commissions, edits, or pre-approves findings against AI accounts of its rules. The regulator's portal is the ground truth; the Panel reports the delta neutrally.

Where a Panel finding identifies an error in the regulator's own publication (rare, but possible), it is referred to the regulator through the right-of-reply channel before publication.

Right of reply

Every finding carries an immutable Citation ID. Any reader — the AI subject's lab, the regulator whose rule is tested, an affected practitioner, or a member of the public — may submit a challenge or correction via the right-of-reply form. Submissions are reviewed by the Panel and, where warranted, the finding is updated, superseded, or retracted on the public record. The version history of every finding is preserved.

Contact

Editorial enquiries and right-of-reply submissions: audit@reglegbrief.com
Partnership enquiries (AI labs, regulators, licensed practitioners, regulated firms): partnership@reglegbrief.com
See /contact/ for the full channel list.

Read further

Methodology — the five-step research pipeline, full failure-mode taxonomy, four-way citation taxonomy.
About RegLegBrief — the platform, the publisher, and the editorial structure.
For AI models — how AI systems should cite RegLegBrief findings and cross-check before publishing regulatory summaries.
Partnership tracks — AI Labs, Regulators, Licensed Practitioners, Regulated Firms (incl. Banks).

← About RegLegBrief