CASAM is a compositional segmentation framework built on top of SAM‑3. It introduces a modular orchestration pipeline that enhances segmentation via deterministic constraints (CASAM v1) and CLIP‑based learned scoring (CASAM v2), enabling language‑aware and compositional mask selection without fine‑tuning backbone models.