RegLegBrief · Verdus Technologies Pte. Ltd.

RLB Panel Speak

The signature publication of the RegLegBrief Specialist Panel. Long-form essays, taxonomies, and arguments on AI hallucinations in regulation. Each piece is a single thought, written for the compliance officer, lawyer, sector team, or AI lab whose work it touches. These are the deep insights and lessons uncovered by the RLB Specialist Panel while working on consultancy assignments for clients.

Latest essay

Published Thursday, 09 July 2026

Excellent formatting, no regulatory credibility

The presentation-substance gap in frontier Claude models — a Panel essay following the three-way and Fable-5 briefings

The three frontier Claude models fail regulatory content differently and land at the same reliability floor. The floor being shared is more important than the failures differing. This essay is on what that gap actually is.

The three frontier Claude models — Sonnet 4.6, Opus 4.7, Fable 5 — land within a three-point band for materially-safe answers on live regulatory questions: 11–13%. That is the reliability floor of the configuration, not a rounding error. This essay argues that bigger, newer, or search-heavier does not close the gap — and that the presentation optimisation is on the wrong side of the tension for regulated use.

By Kratti A Agrawal

essayfrontier-modelsanthropiccompliance-workflowreliability-floormethodology

Read essay →

Excellent formatting, no regulatory credibility

Earlier essays

09 Jul 2026

Fable 5, tested on 102 regulatory questions

How it fails, who it fails, what breaks — the Panel's product-review briefing on Anthropic's Fable 5

Anthropic's Fable 5 commits confidently to 93% of 102 regulatory questions. Only 13% are materially safe. Eighty per cent of confident answers carry factual issues against verbatim regulator text, and 22% of the errors are misattributed citations — nearly 1.5× its frontier siblings. This briefing names five cases of who reads these regulations and what breaks if the workflow trusts Fable's summary.

By Kratti A Agrawal

fable-5product-reviewhallucinationauditcompliance-workflowmethodology

Read essay →

09 Jul 2026

Three frontier Claude models on 102 regulatory questions

A comparative performance briefing — Sonnet 4.6 · Opus 4.7 · Fable 5

An identical exam ran on Claude Sonnet 4.6, Opus 4.7, and Fable 5 — same 102 asymmetric regulatory questions, same instrument, same judge. Only 11–13% of answers per model were materially safe. The three models fail differently, and the failure profiles matter more than the model choice.

By Kratti A Agrawal

comparisonhallucinationauditsonnet-4-6opus-4-7fable-5methodology

Read essay →

19 Jun 2026

The Classifier That Cannot See

When frontier AI models rate independent forensic audit findings on their own hallucinations as low value — a question about whether AI is built to improve, or built to protect itself from scrutiny.

By RLB Specialist Panel

ai-content-classifieradsense-rejectionautomated-moderationplatform-rejectionmeta-ironyaudit-of-the-auditor

Read essay →

14 Jun 2026

The Curse of Recursion: AI Is Eating Itself

And what it means for every regulatory output AI produces

Model collapse is no longer theoretical. A Nature paper named it; the numbers have since confirmed it is already running. As the open web fills with AI-generated content and the rare technical details of regulatory instruments disappear from training distributions, the only durable verification is the primary source.

By RLB Specialist Panel

model collapseAI hallucinationregulatory accuracyprimary source verificationAI safetyprofessional liability

Read essay →

13 Jun 2026

AI Hallucination Is Now a Legal and Regulatory Risk

How courts, regulators and professional bodies across a dozen jurisdictions are placing liability on the professional, not the AI vendor

1,353+ court proceedings globally. $110,000 record sanction in Couvrette v. Wisnovsky. First licence suspension in Nebraska. The classification has shifted: AI hallucination is now personal professional liability.

By RLB Specialist Panel

legal-risksanctionsprofessional-liabilityregulationai-hallucination

Read essay →

13 Jun 2026

Six Types of AI Hallucination in Regulatory Content

Each type has a different mechanism, a different risk, and a different detection method — but they all share one origin

AI hallucinations are not a single failure mode. Six distinct types — H, S, P, SY, E, F — each with a different cause, risk profile, and required detection method. All trace to a training pipeline that learned from secondary sources rather than primary regulator text.

By Kratti A Agrawal

taxonomyhallucination-typestraining-datarlb-hallucination-register

Read essay →

RLB Panel Speak is not press.

It is opinion, analysis, and taxonomies by the RegLegBrief Specialist Panel. For finding briefings and audience-cut releases, see /briefings/ and the Press Room.