This charter states what Aegis Vision does, in the present tense, so that it can be checked. Where a detail is still being confirmed, the page says so. It is revised in place and dated; earlier wording is not quietly replaced.
1. We do not identify people
Weapon and incident detection works on objects and body movement. It does not identify people, match faces against any database, or build profiles. There is one narrow exception, which a school must switch on deliberately. The guard-post module recognises only the enrolled guard's face, only on the camera covering that post, and only when a site enables it. No other person is identified.
2. A person reviews every alert
Every alert is reviewed by a person at the site before any external escalation; the system never contacts emergency services without a confirmed alert except on the highest-severity tier configured by the site. The guard who reviews an alert sees the evidence that triggered it and decides whether to confirm, dismiss or escalate. The system records that decision alongside the alert, so the school can see how its own staff and the software are performing together.
3. Video is processed inside the building
On-premises. Inference runs on a GPU server inside the building; full-resolution video never leaves the site. The guard console runs on the school's local network. The design does not depend on streaming video to a cloud service, and we do not ask schools to accept that as a condition of using the system.
4. The institution controls its deployment
Each school decides which detection modules are enabled, which cameras they apply to, who receives alerts and how escalation is configured. Changes to those settings are made by the school's administrator or by our operator at the school's request, and they are recorded. The system is installed to serve the school's own security response, and the school can withdraw from it.
5. Alert data stays with the school
Alert images are stored on the site's own server. The operator's central dashboard fetches them on demand over an encrypted private network and does not keep copies outside the country. Our operator's view of a site exists to support the deployment, and it is governed by the agreement with the school. We do not pool one school's alert data with another's.
6. Training data leaves only by the school's decision
Only when the school's administrator presses "Send to Hub" are cropped, face-blacked-out, metadata-stripped detection crops transferred for model improvement. Nothing is exported in the background. A school that never presses that button contributes nothing to our training data, and its deployment is supported all the same.
7. We do not sell data
We do not sell, license, rent or trade any video, image, alert record, metadata or derived data from a deployment, and we do not use it for advertising or for any purpose outside the service agreed with the school. Sub-processors, if any are used for a deployment, are named in the agreement with the school. At present no sub-processor has access to video, images or alert records. The encrypted private network used for remote support relies on a coordination service (Tailscale) that cannot read the traffic it coordinates.
8. Adaptation to a site is done by people
Each deployment is fine-tuned on the site's own confirmed alerts; retraining and model updates are started by a human operator, never automatically. Site adaptation uses only alerts that the school's own staff have confirmed, and the results are reviewed before a model update is applied. A school is told when its models are updated.
9. Retention is defined and disclosed
Retention of alert images is set per site and recorded in the deployment agreement; the school decides how long they are kept. Alert records are kept to support the school's own review and incident handling and are deleted when the agreed period ends or when the school asks. At the end of an engagement, the school's data is returned or deleted at the school's instruction, and that deletion is confirmed in writing.
10. We publish accuracy and false-alarm figures once they are measured
We will publish per-class precision and recall, false alarms per camera per day, and the conditions under which those figures were measured, on the transparency page, once they have been measured on held-out footage. Until then the page carries placeholders, and we do not quote figures elsewhere that we would not put there. Behaviour modules currently have no published accuracy figures.
Scope and disclosure
The system is built for schools and universities. Airports; identifying individuals We decline work that falls into those categories.
Security researchers who find a vulnerability in our software or services can report it under our vulnerability disclosure policy, or by writing to security@aegisvision.org. We acknowledge reports, we do not take legal action against good-faith research conducted within the policy, and we credit reporters who wish to be named.
Questions about this charter: privacy@aegisvision.org