SAMBANDH-L3 is a cultural reasoning dataset designed for comparison and elimination-based reasoning over Indian cultural knowledge. Questions require models to distinguish between closely related candidate regions using multiple cultural constraints and relational cues. The dataset includes both English and code-mixed variants spanning across diverse Indian languages, supporting evaluation of contrastive reasoning, constraint satisfaction, and multilingual compositional understanding.