Interpretability Tools as Feedback Loops