Equivocation and ambiguity are common in political interviews, where public figures often avoid directly answering challenging questions. We present our submission to SemEval-2026 Task 6, Subtask 1 on English political response clarity classification. Our system builds on RoBERTa and incorporates NLI-derived semantic features to distinguish Clear Reply, Ambivalent, and Clear Non-Reply responses. To address class imbalance and performance instability, we explore class weighting, multi-seed ensembling, and a hierarchical two-stage framework with threshold tuning. Our best model achieves 60% macro-F1 on the official test set and 64% macro-F1 on an additional evaluation set, demonstrating stable performance across splits. Our results show that carefully engineered smaller models, combined with structured semantic features and imbalance-aware training, provide an effective and computationally efficient solution under limited training data.