Side-by-side animated DAGs contrast observational conditioning P(Y|X=x) with intervention P(Y|do(X=x)). A confounder U creates a backdoor path that biases observation; the do-operator visually cuts incoming arrows into X, leaving only the causal effect X→Y. The visualization also hints at identification via backdoor adjustment when U is observed.
Pure Canvas2D, retro grid-snapped primitives. Two panels crossfade every ~4.2s; animated particles flow along causal edges, with confounder-to-X flow suppressed under intervention. Blocky elbow edges and rectangular arrowheads maintain the homebrew aesthetic; responsive scaling uses scale = min(w,h)/240 with grid size derived from scale.