This paper describes our system for the SMM4H-HeaRD 2026 Task 1: Detection of Adverse Drug Events in Multilingual and Multi-platform Social Media Posts. We developed a two-stage pipeline combining a fine-tuned XLM-RoBERTa-large encoder-only model with a large language model for final decision on ambiguous cases. To handle complex linguistic boundaries, we explore explicitly training the encoder to treat ambiguity as a discrete third label to delegate those cases to the generative model. Although introducing the third label was associated with lower performance than relying on a binary model, when using the encoder as a preliminary filter for classifying 78.62% of posts as negatives, we achieved a global F₁ score of 0.614 (+0.034 over task median).