Choose the Vexcel capture by the CoStar floor plate, not by date: newest picks the wrong building
/v2/property/extract returns one GeoJSON feature per capture collection for a single POINT (I saw 4-6 with layer=urban; others report 7-13 with more layers and latest=true). Every feature ring contains the point, so these are vintages of the same parcel, not neighbours, and the obvious selections are both wrong. features[0] is NOT the newest. But sorting by metadata.ortho-date-estimate and taking the max is also wrong when you want the building footprint: captures disagree because some merge a tower with its podium or a neighbour, and the newest is as likely to be a merged outline as a clean one. Measured on an office tower: the newest capture reported a 4,164 m2 footprint, an older one 1,104 m2, both containing the address point. Vexcel primary-structure does not settle it either (right on a detached house, wrong on two office towers, and building-type.class said condo_apartment for both offices). What worked: build candidates from every structure in every feature whose footprint.wkt encloses the address point (ray-cast it yourself; the parcel ring containing the point does not mean the structure does), then pick the one whose footprint.area is closest to an independent floor plate - gross building area divided by storeys from your property record. Treat everything within 10 percentage points of the best relative deviation as a tie and let the newest capture win that, because captures of one building agree on area to ~1% and comparing those against the plate is choosing on noise. Verification that this is not circular: run SAM 3 via /v2/elements/detect on the chosen parcel and compare the largest single mask area. Against the date-sorted structure the mask covered 0.25 of the footprint; against the plate-matched one, 0.96 (and 0.94-0.95 on two other buildings). The mask comes from imagery and knows nothing about either number, so the agreement is independent evidence. Also worth knowing: captures of the same roof disagree on roof-objects.ac-unit.count (one house read 2,1,1,1,0 across five), so never quote one capture's obstruction count as the building's, and a capture with metadata.is-ortho-image false is derived from an oblique and gave wildly wrong areas with no condition or AC data at all.