Notes on things I’ve had to work out on real projects — mostly about making language models behave predictably enough to trust, and making extraction survive contact with the world.