We present a logical-relations proof of normalization for a language with deep effect handlers and a row-based type-and-effect system. We take the relational interpretation of Biernacki et al. as a starting point and eliminate step-indexing and biorthogonality, resulting in a model resembling a free monad.