spatial capture and geometry on iphone for physical-world apps. the phone scans, and the output has to be good enough to drive real decisions.

so every stage answers one question: is this scan good enough to use.

lidar depth is noisy, meshes drift, coverage has holes. so confidence is an output: depth and mesh agreement gets checked, coverage gets measured, and weak captures get flagged instead of passing through quietly.

nothing downstream should have to guess whether the geometry is trustworthy.

then the question is what to compute on the surface. fitting through noisy real-world data means reasoning about risk, not just best fit, and keeping the heavy work on device so it stays fast and private.

details stay generic here.