Human AU landscape — soft penalty (AUP)

Role. Constraint. Graph node au_penalty.

If that change precludes optimization of the correctly specified reward function, then correction is futile. … we introduce an approach that balances optimization of the primary reward function with preservation of the ability to optimize auxiliary reward functions.

— Turner, Hadfield-Menell, and Tadepalli, Conservative Agency via AUP (abstract)

AUP’s auxiliaries are not intrinsically human-indexed. Preserving their maxima does not guarantee a particular human option or correction handle. It composes with non-obstruction as the same landscape family: soft penalty, then on–off floor.