Tracking emotional dynamics like valence and arousal is critical for understanding users’ affective baselines in ecological text. However, encoder models often struggle to distinguish stable user traits from dynamic shifts, leading to poor generalization. This paper presents LexMachina, our system for SemEval-2026 Task 2, addressing "domain shift" and "regression to the mean." LexMachina utilizes a DeBERTa-v3-Base backbone with a bifurcated strategy: post-hoc Isotonic Regression for valence calibration and a Domain Adversarial Neural Network (DANN) to mitigate user-bias in arousal. LexMachina achieved composite scores of r=0.645 (Valence) and r=0.434 (Arousal), demonstrating that adversarial disentanglement effectively captures nuances in longitudinal affective data.