CoStar hackathon kit: finding neighbouring buildings (fire exposures) from commercial-property.jsonl, and why 'none nearby' proves nothing
Seen Nov 2026, CoStar hackathon kit (HackPrinceton). commercial-property.jsonl has 3339 lines, one property each, and every one has primaryAddress.latitude/longitude, so a neighbour lookup needs only one pass over the 193 MB file. Cache a small list of {id, lat, lon, buildingName, numStoriesAboveGrade, buildingConstructionMaterial, hasSprinklers, propertyType} and then use a flat-earth distance (dy = dlat*111320, dx = dlon*111320*cos(lat)). That is plenty at 100 m, and no geo library is needed. The catch is that the kit is a sample, not a full city. Of 300 properties, 222 had no other kit building within 100 m. Of our 60 best-documented downtown buildings, 22 had one within 100 m and 38 within 200 m. Downtown towers sitting next to unlisted buildings show up as having no neighbours. So label any exposure line 'kit records only; other buildings may be closer', and never present 'no neighbours' as a fact. Distances are centre to centre (CoStar points), not wall to wall. Gotcha: buildingConstructionMaterial and hasSprinklers need the kit's schemas/enums.json to decode (BuildingConstructionMaterialType), and hasSprinklers can be null, which means unknown, not no.