Translation into a pivot language can improve multilingual text classification, but its benefit varies across languages and incurs additional computational cost. We formulate the identification of languages benefiting from translation as a multiple testing problem based on paired classification losses. Within each language, discordant predictions yield a one-sided Bernoulli testing problem, for which a mixture-based e-process quantifies evidence of a reduction in classification risk. The resulting selection procedure controls the family-wise error rate under the stated assumptions. Routing decisions are subsequently fixed and evaluated on held-out data. Applications to SIB-200 and MASSIVE illustrate the trade-offs between classification accuracy and inference latency. The presentation concludes with open questions concerning cost-sensitive selection and extensions from language-level decisions to instance-level routing.
