We study the problem of assessing the clarity of political question–answer pairs, where the goal is to determine whether a response directly addresses the question, avoids it, or remains ambiguous. This task is particularly challenging in political discourse, where evasiveness can be subtle and context-dependent.To tackle this problem, we propose an ensemble-based approach built on the transformer-based model DeBERTa-v3-base, fine-tuned on concatenated question–answer inputs. Special attention is given to class imbalance during training to ensure robust performance across all categories.To better capture uncertainty in borderline cases, we train multiple models with different random seeds and employ Monte Carlo Dropout at inference time. Final predictions are obtained by averaging logits across ensemble models and stochastic forward passes, yielding more stable and robust predictions.Our system achieves a Macro-F1 score of 0.76 on the evaluation dataset. Error analysis reveals that responses that partially engage with the question while failing to provide a direct answer remain the most challenging, highlighting the inherent difficulty of detecting nuanced evasiveness in political communication.