How it knows it's right

Every match passes hard verification gates and a confidence check before it is ever reported.

Every match must pass hard verification gates and a confidence check before it is reported. Confidence is a tier, not a guarantee. See below for what that does and does not catch.

Surface overlap ratio

Point-to-plane registration error

Floor-height agreement

Normal agreement

Catches wrong-wall confusion overlap and error alone can’t see.

Footprint coverage

Catches a thin or partial scan that only covers a sliver of the anchor’s real footprint.

Bidirectional consistency

Registers B-to-A as well as A-to-B on unconfident results and checks the two agree.

Confidence check

Matches are ranked into confidence tiers. Confident results still miss sometimes, which is why we measure and publish that rate instead of hiding it.

Registration has to pass every one of these hard verification gates before an anchor is ever reported as found. Matching uses geometry only: positions and normals, never color or appearance. That is why it works the same way on a HoloLens spatial mesh, an ARKit/ARCore scan, or a depth sensor with no visible light at all.

What the gates catch, and what they don’t: measured, not assumed

Verification rejects most wrong-room attempts: in a 50-attempt adversarial gauntlet against one canonical scan per building, all 50 were correctly rejected. That is not the whole picture, and we are not going to publish only the number that flatters us. Widening the same corpus to every real scan crop we have, not just the one canonical scan per room, turns up confident-wrong results. These are same-apartment, wrong-room matches that clear every shipped gate, plus a rarer cross-building case. These are real. They are in our own regression suite specifically so they can never silently disappear. That is the reason “confidence” below means a tier ordered by how often a result is right, not a promise that it is.

This is the tradeoff we are making on purpose. We publish the failure modes we find, including the ones that make the number worse, rather than quote only the adversarial gauntlet that looks best. Within the correct room, hard borderline cases (thin scans, symmetric geometry) are where the gates are still being calibrated against real captures. A borderline result comes back as “not found” rather than a guess. But “not found” and “correct room, wrong spot with high confidence” are different failure classes. We report both, instead of only the one we’ve solved.

Some geometry has no answer

Some shapes cannot be told apart from geometry alone, by this algorithm or any other. A long, straight hallway with the same width and walls all the way down looks the same from many points along its length. A staircase that repeats the same step shape, or a room lined with identical shelves, looks the same at more than one position. A symmetric room genuinely supports two or more equally good rotations. The room really does look the same from each of them.

Sometimes the shape does not contain enough information to pick one answer. The honest result then is a low-confidence tier or “not found,” never a dressed-up guess. We would rather say “not sure” than hand back an answer that only happens to pass the gates.

See the full pipeline in how it works, or why none of this ever touches a server in offline by design.