Scalable Oversight for Superhuman AI via Recursive Self-Critiquing