Deploying Reinforcement Learning (RL) agents in safety-critical domains remains challenging due to the lack of reliable safety guarantees and interpretable risk estimation. In this work, we propose DESSIE, a Distributional and Explainable Shield inspired by Seldonian policy optimization. DESSIE learns offline an ensemble of safety-aware distributional action-value functions (critics) that estimate the probability distribution of future safety returns, accounting for both aleatoric and epistemic uncertainty. The combination of these distributional critics with neural network ensembles and uncertainty-aware aggregation enables direct probabilistic reasoning over undesirable events such as collisions. The resulting shield activates a recovery policy whenever the estimated probability of unsafe behavior exceeds an interpretable threshold representing the probability of violating a safety property within a future horizon. We evaluate DESSIE on autonomous robot navigation tasks with increasing environmental complexity and compare it against state-of-the-art shielding approaches. Experimental results show that the proposed shielding strategy achieves stronger empirical safety guarantees, fewer unnecessary interventions, and higher task performance.