We propose a percolation-based differentiable d-separation framework that bridges constraint-based causal discovery methods with differentiable DAG structure learning.