<p>This study introduces and empirically validates a comprehensive framework for adaptable, bilingual fake news detection, focusing on English and Arabic. The framework features a custom Transformer architecture that leverages a byte-level tokenizer and multilingual embeddings. Our work presents two primary, empirically backed contributions. First, we demonstrate that the model achieves significantly higher performance on raw, unprocessed text. Notably, this counter-intuitive finding that aggressive cleaning can harm performance was also observed across our traditional baseline models. Second, we validate a continuous learning pipeline using Elastic Weight Consolidation (EWC) over multiple update cycles, experimentally confirming that the model can assimilate new information from evolving data streams while successfully mitigating catastrophic forgetting. While our custom model is outperformed by large pre-trained baselines like XLM-RoBERTa, its value is established as a testbed for these key data-centric findings. The framework’s compatibility with Quantization-Aware Training (QAT) further ensures a path toward efficient deployment. This work provides a validated approach for dynamic, bilingual fake news detection, highlighting the critical importance of both data representation and lifelong learning capabilities.</p>