This poster presents an ongoing study comparing rule-based and deep learning approaches to modeling 13th-century motet composition. The rule-based path uses the harMo13 dataset, symbolic analysis tools (e.g., Music21, Humdrum), and encodes the data in **kern format. Unsupervised learning is applied to extract potential compositional rules. The deep learning path explores fine-tuning, polyphonic adaptation, and custom training. While still in progress, the study proposes a multi-dimensional evaluation framework to assess stylistic validity, rule compliance, and perceptual coherence, aiming to inform both reconstruction and generative modeling of early polyphony.