Abstract
Representation learning methods typically optimize multiple objectives simultaneously, assuming that predictive accuracy, compression, and higher-order organization can emerge in parallel. In this work we investigate an alternative hypothesis: that latent organization develops through a causal sequence of stages. We introduce the MK-VAE framework, which augments variational autoencoders with developmental objectives based on three latent properties: Predictive Depth (PD), Structural Information (SI), and Recursive Topological Return (RTR). These components are combined into a composite developmental metric,