Encoding directional priors as model constraints

This is a first pass: we took a small slice of directional priors from Vastu and Feng Shui, expressed them as structured constraints, and ran a constrained diffusion model against an unconstrained baseline on a 200-plan corpus. The full method, prompts, and notebook are linked at the bottom; this note is the executive summary.

What we did

We picked twelve directional/adjacency rules that appear in both traditions with compatible semantics (e.g., kitchen-in-south-east, no-corner-bed). We encoded them as a small JSON-LD schema with explicit cardinality and tolerance. We then fine-tuned a diffusion-based layout generator with the schema as a conditioning signal and ran a blind preference study with thirty-one human raters.

What we found

• Constrained outputs were preferred 62% vs 38% on overall plausibility (n=240 paired comparisons).
• Light-and-airflow proxies improved 9–14% depending on the room.
• Three of the twelve rules accounted for most of the perceived improvement.
• On the remaining nine rules, the model satisfied the constraint but the preference signal vanished — suggesting either the rule is correct but invisible at this granularity, or it's not actually load-bearing.

What we're cautious about

The corpus is small, the rater pool is small, and "plausibility" is doing a lot of work. We pre-registered the study so we couldn't fish; what landed is what we found. The next pass widens the corpus and looks at downstream outcomes rather than preference.

If you'd like to replicate, the prompts and a smaller open dataset are linked from the board. Counter-examples and methodological corrections are very welcome.