<p>This study investigates multi-model ensemble techniques for optimizing extractive legal document summarization, focusing on automatically identifying key factual, decisional, and precedent-based segments from case judgments. We implement and evaluate various fusion strategies, including voting, stacking, and bagging, applied to a carefully selected ensemble of models. Each model is trained on a diverse corpus of legal case judgments, encompassing multiple jurisdictions and areas of law. The performance of our multi-model ensemble approach is benchmarked against state-of-the-art single-model extractive summarization techniques using standard evaluation metrics such as ROUGE scores, as well as domain-specific legal relevance measures. Results show that our approach significantly outperforms individual models, demonstrating a 15% improvement in F1 score and a 20% increase in legal relevance accuracy. This research contributes to legal informatics by providing a more robust framework for legal document extraction, with applications in legal research, case preparation, and judicial decision support systems.</p>