Ablations for “Frontier Models are Capable of In-context Scheming” — AI Alignment Forum