Methodology

What each label means, how it is decided, and — just as importantly — what it does not claim.

The labels

Satellite

A predicted orbital pass crossed the camera's view cone at the right moment, in the right place, at an apparent angular speed matching what was observed. Position alone is not enough; a satellite that was overhead but moving at the wrong apparent rate is evidence against the match, not a missing detail.

Aircraft

A live ADS-B transponder was where the tracked object was, moving as fast as the object appeared to move. This requires knowing the object's own sky position, which requires knowing where the camera was pointing — without that, "a plane was somewhere in a 93° frame over a city" is a coincidence and no aircraft claim is made. A regular brightness pulse (an anti-collision strobe) can strengthen a match that already stands on position and speed, but never creates one on its own.

Meteor

Fast, brief, straight, and gone. Measured on the duration of the motion rather than the length of the recording — a clip keeps rolling after the object has left, and judging a streak by how long the camera ran rejects every real meteor.

Orb look-alike (Star)

A bright point that stayed essentially fixed against the star field and sat where a known star, planet or the Moon was at that instant. This is a false positive being named. A slightly defocused bright star renders as a glowing ringed disc, which is one of the most commonly reported "orb" sightings anywhere — so it is labelled explicitly instead of being left in the unexplained pile.

Changed course (Maneuver)

Real motion, matching no known explanation, that also broke its own trajectory mid-track. It is separated from Unexplained because it is a categorically different claim: everything else in the sky is ballistic. A satellite is in free fall, a meteor is a rock, and an aircraft turns over kilometres and many seconds. A guard requires the object to have genuinely been travelling before a turn counts — at near-stationary speeds an apparent "turn" is just a centroid wandering a fraction of a pixel.

Unexplained

Real motion in the flagged band that matched no satellite pass, no aircraft, no meteor signature and no celestial object. This is a statement about what was ruled out, not a claim about what the object was. UAP Finder makes no extraterrestrial claim, here or anywhere else.

UAP/UFO — the label the software cannot apply

Every label above is a conclusion the engine reached from evidence: a matched pass, a transponder, a streak, a star at that position, a bend in a path. UAP/UFO is not. It is a claim about what something is, and no part of this pipeline is entitled to make it — so the pipeline is built so that it cannot. The record that carries a classifier's verdict will not hold this label at all; it reaches a sighting only through its owner, after they have watched the clip back.

When someone sets it, the engine's own verdict is kept underneath and shown beside it. That pairing is the whole value: "the person who filmed it calls this a UAP, and the software said Unexplained after clearing 175 catalogue objects and 31 aircraft" is a record worth reading. The word on its own is not, which is why you will never see it here without the machine's verdict attached.

Sky clearance: the negative evidence

Every verdict carries a record of what was checked — how many catalogue objects were propagated to that time and place, how many aircraft were transmitting above the camera, and the separation of the nearest of each. It also records whether each check could be run at all.

That last distinction is the point. A no-match verdict backed by a full catalogue and a sky full of transponders is strong. The same verdict reached because the reference data never downloaded is nearly worthless. Both used to read as the single word "Unexplained"; now they do not.

Confidence values

Confidence is the engine's own estimate in the label it chose, from 0 to 1. It is advisory. Classification never deletes or hides a clip — every detection is kept, logged and reviewable regardless of what the engine concluded.

Bearings and uncertainty

Published sky tracks state whether their azimuths are referenced to true or magnetic north, and carry an uncertainty figure. An uncorrected magnetic bearing is not wrong, but it can carry an offset of up to roughly 20° depending on location, which matters a great deal when triangulating. Where provenance was never recorded, that is published as unknown rather than assumed.

Locations

Exact observer coordinates never leave the server. Published positions are deliberately fuzzed, and the corroboration geometry shown publicly is coarsened to match. This is a privacy floor, not a display preference — a sky camera's location is somebody's home.

Known limits

• Aircraft matching depends on ADS-B coverage. Not every aircraft transmits, and reception varies by region — a no-match is weaker in a poorly covered area, which the clearance record reflects.
• Detection is only as well-aimed as the camera. An uncalibrated mount degrades every geometry-dependent check at once.
• Cloud rejection is deliberately aggressive. A genuine object crossing directly in front of a cloud can be rejected along with the cloud; on a cloudy night, an event you cannot trust is considered worse than an event you do not get.
• Corroboration needs two independent observers with known locations and aims. Most sightings will never have that, and are published unmatched rather than being discarded.