Which Structural Constraints Are Learnable? A Regime Map for a Minecraft Voxel Generator

Neural procedural content generation (PCG) systems produce 3D game content, but practitioners lack guidance on which structural constraints their generator will enforce. We condition a VQ-VAE plus autoregressive transformer with classifier-free guidance on six discretized structural tokens, generate new Minecraft buildings (32³ voxel grids, 513 remapped block tokens), and measure 14 output properties. The properties separate into three regimes: Controllable (9 properties, >100% relative shift; 7 confident), Approachable (4, 20-100%), and Unresponsive (1, <20%). The product of effective signal and training CV correlates with relative responsiveness for emergent properties (Spearman rho=0.879, p=0.002, n=10); the small sample limits predictive generalization. Varying the guidance scale helps diagnose representation ceilings versus frequency floors, giving practitioners a diagnostic before expensive retraining.