Research

The Regla AI Lab

The research arm of Regla ApS. We publish research on how AI-system behaviour is measured, and we hold the AI-native system our own work runs on to the same standard.

Why an audit firm runs a lab

Regla's audits are carried by AI under a certified auditor's signature. That only works if the AI's behaviour is measured, gated, and checked to a standard we would put our name on. The lab is where that standard lives: we study how AI-system behaviour is actually measured, publish the results openly with pre-registered designs, complete transcripts, and reproducibility pins, and run the same discipline over the system our own work runs on.

From the lab to the audit engine

The tools we built to hold our research to scientific standards - pre-registered plans, cited evidence, recomputable numbers, tamper-evident timestamps - are the same tools that gate the audit engine. What we publish is how we work.

Publications

Persona-Name and Grader-Choice as Measurement-Validity Confounds in Agentic-Misalignment Evaluations: A Pre-Registered, Group-Sequential Confirmatory Study

Emil Timmreck Pedersen (Regla ApS) · July 2026 · DOI 10.5281/zenodo.21407064

Does the name an AI agent is given change how it behaves in safety tests? And does the choice of automated grader change the measured result? This pre-registered study tested both on locally run open-weights models, holding everything else fixed. Both mattered: on the primary models, a villain-coded persona name raised the measured blackmail rate, and grading the executed action versus the full response shifted the measured rate by more than 40 percent in relative terms. The effect is model-dependent: a modern aligned model produced zero blackmail across 270 test runs regardless of the name. Every transcript, the grader, and the analysis are public, and every number in the paper can be recomputed from the released artifacts.