Shelter and Access: One-Sided, Falsifiable Interior Measures for Generated Voxel Buildings

Jen-Kai Lin

NeurIPS 2026 Workshop on Embodied Spatial Reasoning (ESR), paper track

TL;DR

Evaluation of open-ended building generation (procedural generators, diffusion models, LLM agents) rests on human ratings, visual models or similarity to a reference. None of them answers: can an agent with a body use this structure as a house?

structgauge answers that with two one-sided measures computed from voxel geometry alone:

measure	question	one-sided claim
shelter	can the building provide cover?	low ⇒ not sheltered (high is not a quality claim)
access	can a player-sized body walk from outdoor to indoor space?	low ⇒ not enterable (high is not a quality claim)

Both are defined on the interior field Ω, which measures how enclosed each air cell is: a discounted multi-bounce field with an exact sparse linear-system solution (no sampling, no learning, no parameter fitted to labels). A low score means the building fails a necessary condition of a usable house (no cover, or no way in), a fact checkable in its geometry. The paper also introduces VoxHouse3D, a dataset of 2,007 player-built houses and 2,023 same-provenance non-house structures, taken from 28 archived worlds of real community servers, one map file each.

Results

Discrimination of house vs. non-house on VoxHouse3D (AUC, same 4,030 builds):

measure	AUC
shelter	0.8681
access	0.8424
size-only baseline (solid fraction, solid_fraction)	0.6550
size-only baseline (log block count, log_block_count)	0.6275

Both measures are one-sided: AUC is reported for discrimination, not as a ranking score. The dataset release recomputes this table from the shipped per-build readings in under a minute (reproduce/reproduce.py, standard library only).

What is in this repository

path	what
structgauge/	the measure implementation
paper/	the accepted paper (PDF)
assets/	figures used on this page
CITATION.cff	how to cite

The dataset (4,030 builds: geometry, labels, per-build readings; CC BY-NC 4.0) and the scripts that recompute the paper's table live on Hugging Face: https://huggingface.co/datasets/jenkai-lin/VoxHouse3D

Quick start

# 1. Install (Python >= 3.9; pulls numpy and scipy >= 1.12)
pip install git+https://github.com/KyleLin0927/structgauge.git

# 2. Measure one build (any .npz in the VoxHouse3D release format, or your own array)
structgauge measure VH3D-w03-a78fe8df.npz

# 3. Run the paper's protocol on a folder of builds
structgauge bench my_generator_outputs/ --label my_generator

Usage

import structgauge as sg

build = sg.load("VH3D-w03-a78fe8df.npz")   # release format
# build = sg.from_array(ids)               # or your own int array, indexed [x, y, z], y up

m = sg.measure(build)
m.shelter, m.access                        # e.g. 0.6526, 0.7641

command	what it prints
structgauge measure BUILD.npz [...]	both readings per build (--json for machine-readable output)
structgauge constants	the constants in effect (discount γ, block semantics, frozen thresholds, block-table version); the defaults reproduce the paper
structgauge bench FOLDER/	alarm rates of every .npz in the folder against the frozen thresholds

bench runs the protocol behind Table 3 of the paper. Point it at a folder of .npz builds from one generator; for every build it computes both measures, then reports the share of builds below the frozen thresholds (the human-house 10th percentile: shelter 0.3748, access 0.4526) as a new row next to the reference rows from the paper (human houses, other structures, and the generators we evaluated).

argument	meaning	default
FOLDER/	folder of .npz builds (release format)	required
--label NAME	name of your row in the table	folder name
--B N	bootstrap resamples for the 95% CI	2000
--json	print the result as JSON instead of a table	off
--thr-shelter X, --thr-access X	override the frozen thresholds; sensitivity checks only, the reference rows are then no longer comparable and the output says so	frozen

What bench prints:

structgauge bench | one-sided alarm rates (frozen human-house 10th-percentile thresholds: shelter 0.3748 / access 0.4526; a rate is the share of builds below the threshold, not a severity)
┌─────────────────────────────────────────────────┬──────┬─────────────────────────┬─────────────────────────┬─────────────────────────┐
│ source                                          │    n │ shelter alarm [95% CI]  │ access alarm [95% CI]   │ either [95% CI]         │
├─────────────────────────────────────────────────┼──────┼─────────────────────────┼─────────────────────────┼─────────────────────────┤
│ * my_generator                                  │   20 │ 0.2500 [0.0500, 0.4500] │ 0.5500 [0.3500, 0.7500] │ 0.6000 [0.4000, 0.8000] │
│ Human house (reference)                         │ 2007 │ 0.1001                  │ 0.1001                  │ 0.1809                  │
│ Same-provenance non-house                       │ 2023 │ 0.7039                  │ 0.6555                  │ 0.8794                  │
│ Creative Agents (image-to-voxel reconstruction) │ 1000 │ 0.9050                  │ 0.6690                  │ 0.9440                  │
│ Build-with-Bombs (voxel diffusion)              │ 2000 │ 0.8865                  │ 0.2050                  │ 0.9220                  │
│ MSG (autoregressive completion)                 │  300 │ 0.8600                  │ 0.4733                  │ 0.9233                  │
│ Direct LLM (pooled)                             │  178 │ 0.2697                  │ 0.5337                  │ 0.7360                  │
└─────────────────────────────────────────────────┴──────┴─────────────────────────┴─────────────────────────┴─────────────────────────┘
frozen: line=(0.37476, 0.45256) = human-house 10th percentile; gamma=0.9; block semantics=v2 (official collision boxes); reference=VoxHouse3D v1.1 (2,007 house / 2,023 other_structure)

The thresholds, γ and the block semantics are frozen in structgauge/report/bench_reference.json, so your row is directly comparable with the paper's.

Takedown and contact

The dataset rests on the work of anonymous community builders. If you are a builder and want a structure removed, open a takedown issue (template provided) or write to kylelin0927@gmail.com; removal ships with the next dataset release.

Acknowledgements

We owe our deepest thanks to the anonymous community builders whose works constitute the dataset, and to the BlockBox community and the UNSW Minecraft Society for archiving their worlds under Creative Commons licences. We also thank the authors of the open generators evaluated in the paper for releasing their systems.

Citation

@inproceedings{structgauge_2026,
  title     = {Shelter and Access: One-Sided, Falsifiable Interior Measures for Generated Voxel Buildings},
  author    = {Lin, Jen-Kai},
  booktitle = {NeurIPS 2026 Workshop on Embodied Spatial Reasoning (ESR)},
  year      = {2026}
}

License

Code: Apache-2.0 (see LICENSE, NOTICE). The dataset is licensed separately (CC BY-NC 4.0) and is not part of this repository.