We describe our submission to SemEval-2026 Task 6: CLARITY, which aims to classify political question–answer pairs by response clarity and evasive technique. We investigate several approaches, including long-context transformers, multiple instance learning, hierarchical multi-task models, and a natural language inference (NLI) formulation. On the development set, our best-performing NLI model achieves a macro-F1 of 0.79 for Subtask 1, while our best attention-based MIL model achieves a macro-F1 of 0.43 for Subtask 2. On the hidden evaluation set, our official submission obtains macro-F1 scores of 0.81 for Subtask 1 and 0.45 for Subtask 2. Our findings demonstrate the benefits of entailment-based modeling for clarity prediction and localized reasoning for evasion detection under limited computational resources.