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

Evaluation of open-ended building generation in embodied environments (whether the builder is a procedural generator, a diffusion model or an LLM agent) still rests on human ratings, on visual models, or on similarity to a reference solution. The first two yield no falsifiable quality claim about a single building, and similarity requires a reference solution that open-ended building does not have. None of them answers the question: can an agent with a body use this structure as a house?

We propose the interior field Ω, defined by voxel geometry alone, which measures how enclosed each air cell is; it has no trainable parameters. On top of it we build two one-sided measures: shelter indicates whether a building can provide cover, and access indicates whether there exists an agent-centric path (one a player-sized body can walk) from relatively outdoor space to relatively indoor space. One-sided means the measure's house-quality claim is made in the low direction only and is falsifiable there: 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; a high score means only that the condition is met, not that the house is good. We also propose 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.

What each measure sees on the same build (VH3D-w07-700de289), on the layer y = 1. (b) Openings on both sides and at the top edge let wind in, so only the middle counts towards shelter. (c) Walkable components for access, one colour per component. Whole-build values: shelter 0.6194, access 0.7726.

On VoxHouse3D, both measures separate houses from same-provenance non-houses far better than building size does.

We then apply the measures to current generators. Each number below is the share of a source's buildings that score lower than 90% of human houses do (shelter 0.3748, access 0.4526). “Either” counts a building that falls below on at least one of the two measures. For every generator, each rate is at least twice the human one. Confidence intervals are in the paper.