Nevermind! I just went down a rabbit hole on "negative hard feature mining" that has simple origins from early days of object detection and bootstrapping!

I'm sure pangram has to deal with a lot of bullshit to keep retraining on new model output though.