Data science competitors design benchmark-style tasks that push coding agents beyond plausible-looking answers. Every task needs its own runnable environment, a checked reference solution and dependable scoring. Contributors run agents on their tasks, study where the agents...