Constraint Learnability Public Release v1.0.2

Patch release for the FDG 2026 constraint-learnability artifact.

Changes since v1.0.1:

• Fix structural-feature extraction to validate manifest-referenced processed files instead of a nonexistent data/processed/builds directory.
• Make scripts/generate.py use the learned any-mask structural prefix when loading a conditioned checkpoint for unconditioned generation.
• Allow scripts/train_ar_conditioned.py to run from existing latent/features files when checkpoints/vqvae/config.json is unavailable, using --num_codes or the paper default of 2048.

Checkpoint assets are unchanged from v1.0.1 and are reattached here for convenience:

• vqvae_step100000.pt SHA256 c9caf9e76a1f1ef8512897cd2aafaa6099d4aa1e7a514cb5203707028e77e019
• ar_cond_step80000.pt SHA256 919b0c1190f940c98cdcd0c58e52c86b3c66702ae8d35b3b7926b2fb8aaec1e7

This release does not change the paper PDF, archived output JSON files, figures, or checkpoint contents.

Constraint Learnability Public Release v1.0.1

Release for the public artifact accompanying Which Structural Constraints Are Learnable? A Regime Map for a Minecraft Voxel Generator.

This release includes the source archive generated from tag v1.0.1 and the final trained checkpoints used by the paper's generation experiments.

Assets:

• vqvae_step100000.pt: final VQ-VAE checkpoint.
• ar_cond_step80000.pt: final conditioned autoregressive Transformer checkpoint.
• SHA256SUMS.txt: SHA256 hashes for the checkpoint assets.

The raw and processed Minecraft structure training data are not redistributed in this repository because the experiments combine third-party datasets, and some upstream sources do not provide clear redistribution licenses for repackaged data. The saved JSON outputs needed to reproduce the paper's tables and figures remain included in the repository under outputs/.