Anyone can record. Finding the exact minute two months later is another matter.

The problem is never too little video: it is too much. Someone describes what happened from memory — “a white van, in the afternoon, Tuesday I think” — and there begins a day lost in front of a timeline. IRIS changes the question: instead of watching hours, you filter. And there are fifteen filters, nameable one by one: date and time, cameras, object type, colour, colour mode, colour tolerance, region of the frame, behaviour, minimum duration, heading, motion inside a drawn area, likeness to a marked object, free text against the video itself, the plain-language sentence and the number of results. They combine with each other, and what comes back is not a stray box but one object with all of its sightings grouped together. When the list is trimmed it says so — so many out of so many — and it never passes an incomplete list off as complete.

The second question comes later: will that video still be there in three months? Each camera has its own retention, and the screen also shows how many days are really on disk — sometimes fewer than configured, because the disk fills up first. On top of retention sit two protections: marking one recording as evidence, and locking a whole stretch of a camera between two times, including what has not been recorded yet inside that stretch. A reason is mandatory and a lock can be left with no expiry. When an incident is opened, the material in its window locks itself. And if the disk fills up and the only old material left is locked, the system stops and writes it out in plain words rather than sacrificing evidence for want of space. Releasing a lock requires a reinforced permission and a written reason, and the record does not vanish: it stays marked as released, with who and why.

Then comes the moment of handing it over. An evidence package is not a loose video in a folder: it carries a manifest with who exported it, when, from which camera, which stretch and under which case reference; a fingerprint for every clip and one for the whole package; a digital signature; and a checker that opens in any browser and works with no connection. That last part is what almost nobody has, and it is what really changes the conversation: whoever receives the video can verify on their own that nobody has touched it — without calling us, without installing anything and without internet. The watermark burns in the real date and time of each frame, not the clock of whoever exported it. The package can be encrypted with a password, and before it is handed over it checks itself: if the check fails, it is not delivered. It is built to stand up in front of a third party; whether it does is always that third party’s call.

All of this connects to the rest of the site at one very concrete point: across the nine zones, every alert keeps its video. This is the part that explains what happens to that video afterwards. From the alert you jump to the exact instant, pull every camera in the wall to that same time and see what happened from several angles at once; the timeline never hides what is missing, and flags each gap with how many seconds it is. We also say what it does not do: likeness search keeps roughly the last two weeks and identifies nobody — it is a resemblance, reviewed by a person; face recognition is available only outside the European Union; and when the AI could not properly look at the video, it says so instead of returning an empty list. Everything travels over a single port, 443, so the network team does not have to open a whole range. And it works with no internet: the models are our own and live inside the customer’s installation, just like the checker inside every exported package.

An incident almost never fits inside one camera. IRIS puts the whole wall on the same instant under a single set of controls: you go forward, back and pause on all four at once, with twelve playback speeds — smooth reverse included — and frame-by-frame stepping when it has to be exact. You jump to a precise date and time, and the link to that instant can be pasted into an email: whoever opens it lands on the same second, not somewhere nearby. When you do not know what you are looking for, a whole day can be skimmed as a strip of stills or as a grid showing when there were people and when there were vehicles, hour by hour, minute by minute or second by second. And what someone writes down stays: a bookmark with its title, its note and its tags outlives the video that prompted it, because a person wrote it.

The question everyone who has ever lost footage asks is where this lives and how long it lasts. Recording spreads across as many servers and sites as needed — different countries and time zones — and is governed from one single view; every clip is stamped with the server that wrote it, the one that serves it and its camera’s time zone, so it plays back and searches the same from any workstation. Retention is set by the customer, camera by camera, and that setting overrides the server default: it can be left with no limit, and an odd value is never read as delete everything. When video does get deleted, it takes with it the eleven kinds of derived data it had generated, each with its own switch in case it is worth keeping; bookmarks are deliberately left out. And none of this is decided at our end: the material lives inside the customer’s installation and goes to no third party.