13.6 Capability and safety composition

A capability adapter for AG News classification and a safety adapter for ToxicChat moderation form a more deployment-like pair. ALLORA reduces pooled hidden order RMSE \(47.7\times \), improves worst-order capability accuracy from \(0.832\) to \(0.851\), and safety balanced accuracy from \(0.800\) to \(0.858\). Toxic recall and benign false-positive rate move from \(0.860/0.260\) to \(0.876/0.172\).

A calibrated weighted additive merge remains stronger on average task scores: capability accuracy \(0.876\), safety balanced accuracy \(0.873\), and toxic recall/FPR \(0.864/0.118\). Its order sensitivity is zero by construction because it is one merged map.