Zoning as a Runtime capability: typed spatial evidence, a deterministic geometry engine, and a fail-safe land-use resolver — over real parcels harvested from official ArcGIS sources.
never ask an LLM a spatial relation a geometry engine can calculate
Every target use evaluated against every harvested parcel by the deterministic-first resolver. The official base zoning is authoritative; anything base-incompatible is PROHIBITED with certainty; and wherever the answer needs an ordinance use-table we have not parsed, it returns UNKNOWN and cites the source — never a fabricated "permitted".
| parcel | zoning | jurisdiction | single‑family | office | retail | warehouse | light‑ind. |
|---|
Point-in-polygon, area and centroid are exact engine computations (CRS-checked, no silent reprojection) — not an LLM guess.
The zoning tenant learns the minimal evidence bundle per situation that still reaches a correct disposition — measured offline vs baselines (examples/zoning_intelligence.py).
Data: parcels harvested by zoning-sources-mcp from official Census/ArcGIS sources (a cross-country sample of the full sweep). Dispositions computed live by the deterministic-first resolver; the geometry engine is dependency-free. Enterprise plugins · architecture.