To trust a model, I need an evaluator to test and interpret how its processes and competencies generate performance across many tasks. As we move toward personalized AI systems adapting to real users, we need models acquiring prosocial competencies coherently and in a way that we can trust. Before considering deployment, models need to align with the principles of logical reasoning that psychologists and philosophers have studied for ages. To start, we can use developmental psychology tasks to verify socially relevant competencies and hopefully slow the pace of capability improvement while prioritizing safety…