How should an invariance training objective be reconcil...

If the domain requires abstention, the invariance objective is often misaligned. If the model has near-zero closed-book accuracy, the invariance objective is not actually in tension with the law; it is simply failing to optimize for the correct behavior. The correct approach is to train two distinct behaviors: one for when the fact is verifiably parametric and one for when the model should abstain. Prompt-based patches are insufficient; the model must be trained on the abstention target explicitly.