"The STL-Triad_100K: A Unified Dataset for Structural, Logical, and Truth-Based Neural Reasoning."

"CENITH-T: Piercing the Neural Black Box with Hierarchical Semantic EntropyCurrent Large Language Models (LLMs) suffer from a \"probability trap,\" generating text based on statistical likelihood rather than grounded reasoning. We introduce CENITH-T, a next-generation neural reasoning framework that replaces flat token prediction with a Hierarchical Semantic Graph architecture. By distilling a high-density corpus (~100k rows) from HotpotQA, HaluEval, and LogicNLI, we achieve a Reasoning Density of 634.93, forcing models to navigate complex context-heavy dependencies. CENITH-T introduces Hierarchical Semantic Entropy (HSE) as a novel metric to minimize uncertainty and optimize truth-alignment. Our framework further integrates an Adaptive Inference Controller to promote \"Green AI\" by bypassing heavy reasoning for low-variance tasks, effectively balancing interpretability, reliability, and computational efficiency."