Why Machine Learning Models Systematically Underestimate Extreme Values II: How to Fix It with LatentNN