Hallucinations in Large Language Models: When More Data Reinforces Error and Minimal Perturbations Reveal Misclassification | Synapse