Water distribution networks face persistent losses of 20–40% of treated water due to leakage, yet conventional detection methods acoustic surveys, hydraulic modeling, and minimum night flow analysis fail to propagate sensor uncertainty through supervisory control loops, while data-driven models lack interpretability and require large, labeled datasets unavailable in under-resourced utilities. This work proposes a modular five-layer supervisory architecture that fuses fuzzy linguistic models, encoding operator expertise and epistemic uncertainty, with Bayesian probabilistic inference for calibrating leak likelihoods from heterogeneous evidence, embedded directly within SCADA data flows. Unlike prior isolated approaches, the framework propagates confidence metrics throughout the decision pipeline, from sensor acquisition to operator dashboards featuring decomposed risk assessments and credible intervals. Validated on a pilot network in Coleipa, Santa Bárbara do Pará, Brazil (360 operational hours, SNR = 12.3 dB), the architecture achieves F1-score = 0.847 ± 0.031 under 5-fold cross-validation, outperforming standalone fuzzy (F1 = 0.762), Bayesian (F1 = 0.791), LSTM (F1 = 0.823), and traditional minimum night flow analysis (F1 = 0.681), while reducing false positives by 34% (95% CI: [27%, 41%], p < 0.01). Under controlled conditions, perfect class separation is achieved (F1 = 1.00). The architecture requires 96% fewer training samples than deep learning baselines and maintains processing latency below 850 ms on networks up to 1,200 nodes. These results indicate that hybrid uncertainty management within SCADA environments offers a viable path toward reliable, interpretable leak detection deployable on commodity hardware in operational water utilities worldwide.