We present a two-stage hybrid system forSemEval-2026 Task 9 on multilingual and mul-tievent online polarization detection. The firststage employs DeBERTa for high-recall binaryfiltering to mitigate severe class imbalance. Thesecond stage leverages Mistral for fine-grainedpolarization classification, enabling improvedsemantic reasoning over candidate instances.This coarse-to-fine design enhances robustnessand efficiency while preserving minority-classperformance. Our system achieves Top-5 results on the English test set, demonstratingthe effectiveness of integrating encoder-basedscreening with LLM-based refinement.