The assistant axis: situating and stabilizing the character of large language models