WFC-JEPA: Constrained World Generation via Wave Function Collapse

Wave Function Collapse (WFC), together with explicit verification, can generate worlds satisfying specified structural constraints, but these constraints do not necessarily capture how a world should evolve. We introduce WFC-JEPA, a modular framework in which WFC generates and verifies possible futures, while a joint-embedding predictive architecture (JEPA) evaluates their agreement with learned dynamics. We develop two complementary workflows: full-trajectory auditing selects complete continuations using predictions across multiple future horizons, whereas stepwise auditing selects successive next states using predictions conditioned on the accepted history. Both operate without access to the true future, and support deterministic or probabilistic evaluation. In a controlled corridor experiment, multi-horizon JEPA improves selection over single-horizon evaluation, and outperforms an incomplete handcrafted motion model. Larger candidate pools improve accuracy, although physically realizable continuations remain scarce. In a second experiment with two plausible futures, probabilistic evaluation improves stepwise selection, while mode-locking prevents switching between routes. These findings demonstrate the potential of combining explicit structural constraints with learned predictive guidance, also highlight the limitations of this combined candidate generation and predictive selection method.