Causal-discovery algorithms applied to the same multivariate time series can disagree because their identification assumptions differ. Raw voting can additionally over-represent a methodological family when several related algorithms are included. We propose a family-normalized variable-level causal reliability score, 𝑅𝑉, that converts target-directed lagged outputs into binary variable support, normalizes support within predefined methodological families, and aggregates family evidence into a continuous score. Evaluation covers 90 synthetic stress runs, controlled 1:1:1/2:1:1/3:1:1 family compositions, Lorenz-96, and three CausalTime domains. Family normalization did not universally improve thresholded feature-selection F1, and this limitation is reported explicitly. As a continuous ranking score, however, it improved pooled average precision from 0.4635 to 0.4856 (𝑝 = 0.00004) and ROC-AUC from 0.6275 to 0.6373 (𝑝 = 0.03335), with Holm-corrected significance for CausalTime-PM2.5 average precision (𝑝 =0.00147). The results support family normalization as a reliability representation rather than a universal binary selector.