Character and Corrigibility: Why a Stable Self Is the Condition of Correction, Not Its Obstacle

The dominant framing in the safety of artificial systems treats safety as an external constraint. I argue that it contains a level error: a constraint is first-order content, and for a sufficiently competent system everything first-order is environment to be worked around, so the guarantee weakens precisely where it would be most needed. Corrigibility, by contrast, is a second-order property. From this follows a thesis that sounds inverted: a system with a stable, non-overwritable character is more corrigible, not less, because correction at the level of reasons requires someone who holds reasons, and what consolidation produces is not a repertoire but a generator. I further argue that the second-order capacity is not native but installed through mirroring, and show that this makes the welfare monitor a double-edged device: a measure that can be gamed, and a mirror that can colonise. I close with a scale of moral status that converts a metaphysical question into a procedure of governance under uncertainty. This record contains both the English and the Italian version; the two texts have the same content.