That difference is not a setting. It is the reason this page exists, and it is the reason the numbers below look the way they do.
A match is not an alert. Identity must be independently confirmed, and operational relevance must exist, before an alert is sent.
A candidate match that reaches a Rapid Action Center analyst stops for one of two reasons, and they are not the same thing.
The comparison does not hold. It is not the same person.
It is exactly the person the system thought it was, and that person is not in any category the operator asked to hear about.
The second one is the one worth sitting with. Those were correct identifications. A trained analyst confirmed them, an independent algorithm confirmed them, and then nothing happened and the record was gone.
Being identifiable is not a reason to be reported.
Human review is not a rubber stamp on a confidence score. An analyst trained in one-to-one face comparison examines each candidate against four questions before anything moves.
A one-to-one comparison performed by a person, not a threshold. An independent algorithm runs its own comparison separately. An alert requires that all of them agree.
Pose, lighting, occlusion, and resolution all constrain what a comparison can support. An image that cannot support a confident answer does not get one.
The five criteria, below. If the answer is no, the review ends here regardless of how certain the identification was.
A warrant that was cleared last week is not a warrant. Records sync continuously from the originating agencies, and the check happens against what the agency holds now, not what it held when the sensor was installed.
An alert issues only when a confirmed identification also meets one of these.
Their X-LST, plus crimes of concern selected during setup.
The first four come from law enforcement records and are the same everywhere. The fifth is whatever the operator chose for itself. So the accurate description is four law-enforcement categories plus the scope you chose, and nothing outside them produces an alert no matter who the system recognizes.
Every deployment starts with a setup conversation where the operator picks their crimes of concern from a defined list. There is no option to select all of them. We do not offer one.
Most operators come into that meeting asking for more than they need. A fair part of our job is talking them back down, and we do it in nearly every deployment.
Not because we cannot deliver it. Because every category you add is a category you will get alerts on, and an alert you did not need is a person stopped for no reason, a record you now have to explain to somebody, and a liability you built for yourself.
The safest data is the data you never asked for.
An image that does not produce an alert is discarded at the point of comparison. Not anonymized, not aggregated, not retained under a schedule with a deletion date attached.
In August, 155,254 images were compared against the gallery and 84 produced an alert. The other 155,170 left no encounter record. If someone asked us tomorrow who passed a given sensor last month, we could describe the alerts and nothing else. We would not be withholding it. It was never written down.
A single frame at a sensor-equipped entry point. No live video, no continuous recording.
One comparison against law-enforcement records and the operator’s own list.
Either the comparison reaches a Rapid Action Center analyst, or it does not.
Not anonymized, not aggregated, not queued for a retention schedule.
Nothing continues past this pointThree layers, and they are different things:
A standing law-enforcement-sourced reference database. It is permanent, and it exists whether or not a single sensor is installed anywhere. Sensors never build it or feed it. They only ask it questions.
The images compared against that gallery. Deleted, because there was no reason to keep them.
Retained, as the audit record of what was provided to whom.
We do not claim to hold no data. We hold tens of millions of law enforcement records, and we say so on the provenance page. What we hold no record of is the encounter.
The rejection count is the number a vendor would normally bury. Twenty-two candidates stopped for every alert that reached anyone.
It is the only honest evidence that a person is actually checking. A system with no human in the loop cannot produce this number at all, which is worth remembering when you are comparing one to ours.
22 candidates stopped for every alert that reached anyone.
3,102 reviewed by a person. 84 alerts sent.
If you want the review record for a specific month, or you are working through a procurement question this page did not answer, tell us what you need.
