Structured tasks with clear input-output specifications where model reasoning can be automatically checked for correctness.