TINF-Conformal Predictive Sets: Separating Neutrality from Ensemble Indeterminacy with Finite-Sample Coverage

Probabilistic classifiers can be uncertain because each plausible model is itself ambivalent, or because individually decisive models disagree. A mean posterior and even its entropy cannot identify which mechanism is present. This paper gives the four-component (T, I, N, F) neutrosophic distinction between pure indeterminacy and neutrality an operational, testable form for ensemble classification. For candidate class y, truth is the ensemble mean probability, falsehood is the largest rival probability, neutrality is mean within-member binary entropy, and indeterminacy is the Jensen gap between entropy of the mean and mean entropy. The construction is bounded, permutation invariant, and satisfies the exact identity N_y + I_y = h2(T_y). It distinguishes two ensembles that have the same mean and total entropy: unanimous 0.5 predictions produce pure neutrality, whereas an equal split of deterministic 0 and 1 predictions produces pure indeterminacy. We prove this non-identifiability result, finite-committee concentration bounds, and a distribution-free split-conformal coverage theorem for a monotone TINF nonconformity family. Experiments use 50 independent random partitions of four public benchmark datasets. These findings support TINF profiles as interpretable diagnostics wrapped by conformal validity.