This work analyzes two forms of optimization bias that leave 3D Gaussian Splatting reconstructions blurry. Random Seeding and Random Splitting introduce explicit exploration to address these limitations and recover scene detail. The two operators target distinct limitations in gradient-based optimization and are evaluated separately and together across scene-reconstruction datasets.