Our work spans mechanistic interpretability, activation steering, and evaluations of AI systems in high-stakes environments such as legal reasoning, live production software, and civic institutions.

We publish openly and build evaluations and tools the broader research community can use.

The behaviors that matter most for safety, like planning, patience, and strategic waiting, only show up over a horizon. Reading and steering them in current models is an open problem, one that sits underneath deception, reward hacking, and agent control.

We study how models encode time and intertemporal preference, how those representations can be steered, and how AI systems behave in high-stakes environments.