to be fair the constitutional model is a winner, claude actually pays a lot of attention to its context and for my current work that means ignoring all the nonsense from training data that contradicts the way i want it to work. but that's pretty relative, it still is constantly falling into what i call the "gravity well of training data".

i am going to look into building harness and using open models soon though. partly for my own work and partly to maybe get a job doing this work.