Data sources and geometry contract

The pipeline downloads a bounded subset of public road and parcel geometry. It records the URL, acquisition time, canonical JSON SHA-256, field whitelist, feature count, and requested bounds in data/raw/sources.json. Raw and generated geometry are local artifacts, excluded from Git until redistribution terms are reviewed.

Input	Source	Current use
Parcels	HCAD parcels served by Harris County	Geometry, OBJECTID, and state_class only
Roads	City of Houston street centerlines	Geometry, OBJECTID, NAME, and ST_TYPE
Ground elevation	USGS 3DEP elevation service	Houston 2024 catalog raster, NAVD88; observed DEM acquired
Historical canopy footprint	H-GAC Urban Forestry CIR layer 3	Service labeled 2016; acquired footprint, no stems/heights/species
District vegetation heights	USGS 2018 Coastal LiDAR EPT	District returns extracted and rendered; stems/species unverified. See evidence.

No Google Earth imagery is downloaded or redistributed. Regional land-cover forest classes cannot establish individual live-oak placement. The catalog does not imply that missing datasets have been licensed or acquired.

Acquire and build

Edit the provisional WGS84 bbox in config/river-oaks.json before acquisition. The current rectangle extends beyond the requested neighborhood. An authoritative boundary polygon is still needed before claiming full River Oaks coverage or the brief's approximate acreage.

ArcGIS object IDs are enumerated first. Feature chunks use POST to avoid the county server's GET URL-length limit. Each chunk must return exactly the requested IDs. Only allowlisted attributes survive acquisition; owner/mail records never enter the local artifact. Features are clipped in a projected meter CRS, and invalid geometry aborts the build.

world.json uses schema_version: 1, crs: EPSG:32615, and horizontal local coordinates relative to the projected origin. X points east, Y north, Z up. Unreal converts (x,y,z) to (100x,-100y,100z). With --terrain, Z is absolute NAVD88 elevation in meters; without it, the fixture is flat at zero and cannot pass elevation checks.

• roads: stable source-part ID, name, polyline points, provisional width of 8 m, unknown lane count. Widths, medians, signals, and esplanades require observations.
• parcels: stable source-part ID, exterior ring, and interior holes. Geometry keeps its original shape.
• buildings: parcel ID, ground-centered location, width/depth/height, yaw, and seeded style. Only A1 residential parcels receive house massing. Footprints shrink until the source parcel inset contains them. Style/height choices are synthetic and all houses enter manual review.
• trees: observed location, crown radius, height, species label, and source ID. No input means no generated trees.
• provenance: source hashes and acquisition receipts. Generated counts are not accuracy evidence.

Observed ground and canopy mask

river-oaks observe locks the USGS export to catalog raster TX_Houston_B24 instead of silently selecting a different mosaic. The source catalog describes 1 m native resolution; the bounded 513×513 download is about 8.82 m/pixel. The browser mesh resamples that export to 257×257. The source receipt records the catalog raster, export parameters, returned dimensions, SHA-256, CRS, and NAVD88 datum. The catalog date is not independently verified as the flight date.

Road segments are densified to at most 20 m before sampling elevations; house and tree bases sample the same DEM. Bilinear sampling fails on missing or out-of-bounds pixels. verify --terrain independently resamples every grid vertex and feature height; this tests import fidelity, not survey accuracy. The optional manifest terrain uses south-to-north rows and west-to-east columns. grid_origin_m, spacing_m, width, height, and flat heights_m specify that grid.

The H-GAC export preserves its returned extent, which may be wider than the requested bbox. Only the verified tree-canopy color is classified; unexpected renderer colors fail acquisition. canopy-reference.json records the resulting local-meter Polygon/MultiPolygon, export resolution (about 4.32 m/pixel), source label year, and receipt. Native imagery resolution and exact acquisition dates are unknown. The service exposes no clear redistribution license, so downloads remain ignored local inspection artifacts. This overlay is a historical reference and does not populate trees. A later LiDAR crown extraction must use an independently sourced observation and account for date/leaf-season differences.

Supply canopy observations

Convert licensed LiDAR-derived individual trees to a WGS84 GeoJSON FeatureCollection. Add a top-level source_id, a unique feature id, point geometry, and properties crown_radius_m, height_m, and optional species. Unknown species should remain unknown.

uv run river-oaks build --trees data/raw/observed-trees.geojson \
  --output unreal/Content/Data/world.json

For independent canopy verification, supply a separate reference file with source_id, crs: "local_m", and a Polygon/MultiPolygon geometry transformed into the manifest's local frame. Reference provenance must identify a distinct measurement source from tree placements. Do not simply rename the same layer. The verifier rejects identical source IDs, but establishing source independence also requires reviewing source receipts.

uv run river-oaks verify --world unreal/Content/Data/world.json \
  --canopy data/raw/independent-canopy.json

The verifier compares crown coverage fraction and intersection-over-union, so equal canopy area in the wrong location fails. IoU ≥ 0.75 and coverage error ≤ 0.05 are provisional thresholds that require agreement for the selected sensor resolution.

Interpret verification

Road and parcel checks compare generated geometry against independently reconstructed source geometry, including missing/extra IDs. The road threshold is a strict maximum Hausdorff deviation below 5 m. A zero import deviation proves source geometry survived processing; it does not establish source survey accuracy. Building checks independently reconstruct rotated footprints and require containment in source parcel setbacks. Bounds must match the projected clipping extent.

Release gates remain blocked for independent survey accuracy, exact neighborhood boundary, terrain, lane topology, residence similarity, production visuals, live Jev, and GPU performance. The verifier has no switch to override these missing gates. Add a verified evidence importer with each production milestone instead of treating a manual boolean as proof.

residence-review.json queues every synthetic house for review. No facade similarity detector or licensed comparison corpus exists yet. There are no recreated private interiors or copied architectural plans in this project.