Testing for Scheming with Model Deletion — LessWrong