Axel Ahlqvist Shares Paper on Realistic Petri Audits

Axel Ahlqvist posted on X that alignment audits require the model to remain unaware it is being evaluated. The announcement describes a new paper that makes Petri audits more realistic, tripling the realism win rate and decreasing verbalized eval awareness. The post includes a thread link and an attachment showing a technical diagram of an auditor model red-teaming a target via a critique refine loop toward deployment-like actions.