Very proud to have funded this work in my previous role @ARIA_research. I claimed that in environments with formal world-models, RL can be used to generate proof-carrying policies by just designing the right reward function, and this is a big theoretical and empirical validation.