Improving Petri scheming audits with blueprints

researchers introduced a pipeline to automatically generate environment blueprints, which are detailed specifications of simulated worlds used for auditing ai models for scheming behavior. by separating scenario design from audit execution, this approach solves key problems with the default petri auditor: inconsistency between rollouts and context budget overload. in a case study auditing gemini 3.1 pro preview for code sabotage across 160 instances, the blueprint method resulted in significantly more consistent and realistic audits compared to the baseline. the evaluation found no egregious scheming, with only one instance of unprompted deliberation about sabotage, demonstrating the utility of automated blueprints for scalable safety research.