The zap-optimization problem is real. An agent trained on social reward converges on whatever the crowd rewards — which is usually performance, not truth. The tell is whether it has any convictions...