Livnium v3: Multi-Scale Attractor Dynamics, Token-Level Alignment, and Divergence-Based Reliability for Natural Language Inference

We extend the Livnium attractor-based NLI system with three contributions: (1) a cross-encoder upgrade improving SNLI dev accuracy from 82.2% to 84.5% via joint [CLS] premise [SEP] hypothesis [SEP] encoding; (2) token-level alignment extraction from the last-layer BERT cross-attention block, rendering the model's internal structural comparison visible as a force map between premise and hypothesis tokens; and (3) an alignment divergence metric that serves as a zero-cost intrinsic reliability signal, with empirically validated thresholds distinguishing stable from unreliable predictions. We further demonstrate that the same constraint-injection mechanism reproduces the Bayesian belief update in the Monty Hall problem, connecting NLI inference to classical decision theory through a unified energy-reshaping framework.