Promises to “pace” advanced AI mean little without clear limits. The real test is whether safety oversight can stop a dangerous project—not simply monitor it.