False positives and false negatives

False positives: a match that is not smart glasses

Shared company identifiers Known fact

A Bluetooth company identifier is registered to a company, not to a product. Meta Platforms Technologies, LLC holds 0x058E; the same entity ships Quest headsets and controllers. Xiaomi's identifier covers an enormous range of consumer electronics. Luxottica's identifier could in principle appear on any connected eyewear it makes. An advertisement carrying one of these identifiers tells you the manufacturer's family, nothing narrower.

How the model contains this Known fact

The confidence engine caps any match made up of a single kind of evidence at 0.49, below the MEDIUM threshold, no matter how many times the device is seen. Company-identifier-only matches therefore stay LOW and are never surfaced as consumer notifications. Reaching MEDIUM requires a second, independent kind of evidence, such as a service UUID or a payload pattern, and reaching HIGH requires a third.

The open question Unknown

Service UUID 0xFD5F is registered to Meta Platforms Technologies. If Quest controllers also advertise it, then the company identifier plus service UUID combination would produce MEDIUM detections near a Quest, which is exactly the false positive the cap is meant to prevent. This cannot be reasoned away; it has to be measured. Block D of the validation protocol tests a Quest and its controllers, a busy café, and an ordinary day, and records every MEDIUM or HIGH that fires without glasses present.

Other sources of false positives

Byte-pattern collisions. A pattern like an ASCII string in the manufacturer payload is only evaluated when the company identifier matches the signature's. The same bytes under a different company identifier are ignored.
Name collisions. Local names can be changed by users on some devices and copied by others. Names are weighted evidence, not proof, and are only available in the foreground for devices that put them in the scan response.
Demo mode. The app's demo uses a fictional signature under 0xFFFF, a value the Bluetooth SIG reserves, so a demo detection can never be confused with a real one and is labelled DEMO in purple throughout.

False negatives: smart glasses that are present but not detected

Silence while connected Inference

Many Bluetooth accessories stop advertising once they are connected to their host device, because advertising exists to enable a connection. Whether the glasses we study behave this way while worn and connected to their owner's phone is unknown until tested, and it is the single most important question in the validation plan. If they are silent, a scanner cannot see them, and no scoring model can change that. This is a physics and firmware limitation, and the product will state it plainly.

iOS background limits Known fact

Apple documents that background scanning requires an explicit service-UUID filter, coalesces repeated discoveries into one event and runs at a longer interval. A device that advertises without a service UUID cannot be discovered by a backgrounded app at all. Read what iOS lets a Bluetooth scanner do.

Address rotation and deduplication Known fact

iOS does not expose hardware addresses, and the peripheral identifier it provides can change when a device rotates its private address. The app treats a rotation as a new session rather than trying to link sessions, so a long-present device can appear as several short sightings. That is a deliberate privacy choice.

Environment

Bodies, walls, bags and pockets attenuate 2.4 GHz signals. A device across a room may fall below the receiver's sensitivity in one position and not in another. Block B of the protocol measures this at fixed distances in a quiet room and in a busy space.

How the two error types will be reported

When results exist they will be reported per scenario, not as a single accuracy figure, because a single figure would hide exactly the trade-off that matters: a threshold that avoids Quest false positives may miss glasses in some states, and vice versa. The calibration rule is written down in advance: the minimum evidence set seen in at least 80% of glasses-present scenarios must yield MEDIUM, and any evidence set that also appears near non-glasses hardware from the same manufacturer must yield at most LOW.

How Bluetooth detection works, including the scoring formula
Where signatures come from
RSSI, proximity and distance
Device database