Context-Regulated Q-Neutrosophic Soft Model Selection under Input Perturbation: Evidence from Three Classification Benchmarks

Selecting a classifier using a single average metric can conceal instability across validation splits, poor lower-tail behavior, and sensitivity to input perturbations. This study formulates a context-regulated Q-neutrosophic soft model-selection procedure with two fixed evaluation conditions. For each classifier and performance criterion, mean attainment (truth), across-split dispersion (indeterminacy), and the complement of a lower empirical quantile (falsity) form an auditable three-component descriptor. A bounded lower-context regulator aggregates their penalized utilities; monotonicity and perturbation bounds follow directly from its convex representation. Four fixed classifiers are compared using five benefit-oriented criteria, three repetitions of stratified five-fold cross-validation, and paired test-feature perturbations. Three publicly distributed benchmark datasets cover breast-cancer diagnostic features (569 records), wine chemistry (178 records), and handwritten-digit images (1,797 records). Under the stated weights and penalties, the radial-basis-function support vector machine ranks first on all three tasks, with composite scores 0.961179, 0.981183, and 0.979696, respectively. The difference from logistic regression on the diagnostic benchmark is only 0.000349 and reverses in some sensitivity checks. Comparisons with unpenalized aggregation, worst-context means, variability penalties, and ideal-distance ranking show that the proposed score supplies explicit uncertainty diagnostics, not demonstrated universal predictive superiority. The complete calculations and assessment-level outputs are provided for independent reproduction.