Logical Counterfactuals Consistent Under Self-Modification — AI Alignment Forum