Demonstrating specification gaming in reasoning models — LessWrong