There's a question the alignment field keeps asking: How do we make models better at monitoring themselves?