When conditioning a generative model on structural properties, some constraints respond immediately while others barely change. We measure 14 structural properties under conditioning interventions in a VQ-VAE + autoregressive transformer pipeline trained on Minecraft buildings (32³ voxel grids, ~260 block types) with classifier-free guidance. The properties separate into three regimes: Controllable (7 properties, >100% shift), Approachable (5), and Unresponsive (2, <20%). Two pre-experiment features, effective signal strength and training variation, predict regime membership with 83.3% leave-one-out accuracy (ρ=+0.661, p=0.038). Varying the guidance scale at inference shows that Unresponsive properties are bounded by the decoder's representational capacity, while some Approachable properties are bounded by data frequency. The regime map and predictor help identify whether a given constraint requires architectural changes or more training data.
Alex Li (Fri,) studied this question.