I want to lay out why I think detection work is structurally harder than generation work, because I keep seeing people compare them as if they are symmetric problems. A generator only has to be convincing once. A detector has to be right continuously, against an adversary who gets to see the detector's output and iterate against it. That asymmetry compounds. Every published detection benchmark becomes a training target the moment it is published. So when I look at something like BitMind the question I care about is not the current accuracy number. It is whether the incentive structure keeps producing new detection approaches faster than the generators adapt. That is a much harder thing to evaluate and I do not think anyone has a clean answer yet.