SEMTRA: Global Semantic Transition and Rough-Set Rules for Auditable Post-hoc Explainability

Deep learning architectures produce highly effective yet uninterpretable latent representations, creating an interpretability gap that fundamentally impairs model verification and automated rule extraction. In this work, we propose Global Semantic Transition (SEMTRA), a post-hoc framework bridging this gap by translating numerical representations into auditable rough-set production rules without retraining the underlying feature extractor. Evaluated on the Animals with Attributes 2 (AwA2) benchmark, the proposed semantic transition achieved a mean absolute error of 0.1295. The systematically extracted rulebook successfully covered 86.40% of the test instances, significantly outperforming standard separate-and-conquer learners, which covered only 26.20%. The rules yielded a non-abstained covered accuracy of 40.73%, strictly verifying the transparent, mathematically robust portion of the model logic. Furthermore, the continuous semantic prototype transfer achieved a zero-shot accuracy of 48.43%, surpassing foundational baselines such as direct attribute prediction, which achieved 46.10%. A synthetic benchmark independently confirmed exact algorithmic recovery with a macro-F1 score of 0.8668. These findings establish that post-hoc semantic transition, combined with rigorous rough-set granulation, provides a reproducible pathway to transparent symbolic explanations, effectively balancing the tradeoff between predictive precision and algorithmic verifiability.