Capturing both transitional physics and long-horizon behavior in the latent representations of learned dynamical systems remains a central challenge of world modeling. Existing approaches typically rely on explicit latent structures or strict short-term inference, both of which limit the ability to represent emergence and support open-ended discovery.

We explore a discovery engine composed of a priorless generative model and a diffusion-based sampling appendage that identifies novelty within its associated latent space. Using Conway's Game of Life as our environment, we train our two-headed generative model on the gradient between adherence to B3/S23 physics and the clustering of latent trajectories according to five key behavioral signals (ΔP, Δcx, Δcy, V, D). Using novel latent distancing from known behavioral centroids, alongside remaining on the frontier of generational feasibility, is what brings scalability to our emergence.

Latent collapse, rollout drift, non-stationary novelty signals, and compound normalization are all addressed by our staged curriculum, contrastive regularization, and dynamic clustered-novelty libraries. This approach not only generates rich, navigable, and scalable latent cultures but also opens the door to future self-probation and emergent applications beyond the GoL plane.