Ensemble learning improves predictive performance by combining multiple classifiers; however, indiscriminate aggregation introduces redundancy, computational cost, and limited interpretability. This paper proposes a compact stacked ensemble framework that selectively retains classifiers exhibiting reliable probabilistic behavior. A γ′-pruning model based on fuzzy-rough set theory evaluates classifier relevance using out-of-fold posterior probability estimates, Gaussian similarity relations, and a fuzzy-rough dependency measure to capture probabilistic consistency beyond accuracy-based criteria. A greedy selection strategy balances validation performance with probabilistic stability. Experiments on medical and bioinformatics benchmarks demonstrate significant ensemble size reduction without sacrificing predictive performance, supporting fuzzy-rough probabilistic relevance as an effective criterion for classifier selection in uncertainty-aware stacked ensemble systems.