Resolution of learned-direction collateral-damage differences

Resolve whether learned direction constructions differ from random directions in semantic-neighbor and capability damage under the random-calibrated intervention-path evaluation.

Background

The study finds that learned directions improve target leverage and clean-path incidence relative to random controls, but the corresponding differences in semantic-neighbor and capability damage are not statistically resolved. The unresolved issue is whether learned localization and geometry methods produce a systematic collateral-damage advantage, disadvantage, or no difference relative to random intervention directions.