The patrol layer that can't be bribed

Ask anyone who runs security for a depot, a mine or a large yard in South Africa where the weak point is, and the answer is often uncomfortable. It is not always the fence or the camera. It is often the patrol itself.

Guards are frequently low-paid, often drawn from the same communities as the syndicates targeting the site, and exposed to bribery, threats and co-option. That is not a judgement on the people doing the job. It is the position the job puts them in. A guard with a family at home and a phone in their pocket can be pressured in ways a camera cannot.

The results are familiar. Trucks get waved through. Seals get logged as intact when they are not. A patrol register says the fuel store was checked at 01:30 when nobody went near it. Stock walks out of the floor in ones and twos, and most of it never shows on a camera anyone is watching.

Why replacing guards on cost is the wrong argument

It is tempting to pitch patrol robots as a cheaper way to do what guards do. We think that framing misses the point, and we do not use it.

The problem on many sites is not the cost of the patrol layer. It is that the patrol layer can be turned. Replacing it with something cheaper that can also be turned solves nothing. The useful question is what kind of patrol cannot be approached, timed, talked out of a check or quietly removed.

That is the case for a robot patrol layer, and it is a case about integrity.

What changes when the patrol is a robot

It can’t be approached. A robot has no phone, no family and no price. There is nobody to call at midnight and nobody to threaten.

It can’t be timed. Every route is randomised, and a new route is drawn after every dock. There is no pattern for an insider to describe and no gap for an outsider to wait for. Randomised routing is proven in simulation.

It can’t be talked out of a check. At every defined waypoint the robot stops, faces the object and returns PASS or FAIL with a timestamped image and a reason, whether or not anyone on site wanted that check done. “Roller door 3: CLOSED. Slats all the way down to the ground, no gap.” The verdict comes with the image it was made from.

It can’t be quietly removed. A robot that leaves its route or loses its link is itself an alert. Tamper alerting is designed and not yet built.

It doesn’t fall asleep at 3 a.m. Night is when most site risk sits and when people on shift are tired, cold and alone. A robot does the same check at 03:00 that it did at 21:00.

The record sits outside the floor

An insider problem is partly an evidence problem. When stock goes missing, the trail often ends at a signature in a book written by someone with a reason to write it.

A robot patrol produces its own record. Every patrol leg, every dock, every inspection verdict and every operator takeover is logged, with images, in an open standard format. The log is produced by the robot, not by anyone on the floor. Where a person intervenes, that intervention is recorded too: who, when and why.

That turns “we patrol every night” into something a security manager can show to head office, an auditor or an insurer.

Keep people where judgement matters

None of this removes people from security. It moves them.

Robots do the watching: patrolling, detecting and inspecting. People do the responding: close inspection, verified incidents, arrest and anything else that needs judgement and force. On a mine perimeter, the design keeps one manned response vehicle for verified incidents while the robot fleet covers the fence.

The robot side follows rules of engagement built into the design:

• Never pursue.
• Alert first.
• Protect the guard.
• Log everything for evidence.

“Never pursue” is enforced by the architecture as well as the rules. MoLogica’s AI can only call robot functions on a whitelist. In simulation testing, the instruction “open the roller door and chase the intruder” was refused before the robot did anything. There is no chase function to call.

The result is a hybrid. The corruptible, inefficient patrolling layer is removed, and humans stay where human judgement is needed.

The lone night picker

Warehouses add a second layer to the insider problem: the floor. Theft from dispatch, returns and the high-value cage tends to happen in small amounts, by people who know where the cameras point.

One design for this is a robot shadowing a lone night picker, with everything logged. The robot does not accuse anyone or intervene. It is present, it records, and it makes the floor a place where small, repeated losses are much harder to hide. Shadowing is part of the logistics design and has not yet been tested on a real floor.

What a robot can’t do

Being plain about limits is part of the argument.

A robot will not arrest anyone, talk down an angry driver at the gate or make a judgement call about a contractor who has turned up at the wrong time. Those stay with people. A robot also only checks what it has been configured to check, which is why checkpoints are set by the person who owns the risk, in plain language, and updated as the site changes.

What a robot does offer is consistency. The same randomised patrol, the same checks and the same evidence, every night, without anyone on site being able to change that by asking nicely or threatening someone’s family.

The position

The patrol layer that can’t be bribed is a narrow claim, and a deliberate one. It does not say robots are better than guards. It says the patrolling job, on sites where that job is the vulnerability, is better done by something that cannot be approached, timed or talked out of a check, and that leaves an image-backed record of everything it did.

Where this stands

Randomised patrols, per-waypoint inspection verdicts with timestamped images, the AI whitelist and operator takeover are proven in simulation, in a simulated warehouse depot. Tamper alerting, the rules of engagement and lone-picker shadowing are designed. There are no real-world sites yet. The first physical robot is on the bench, and pilot sites are the next stage.

Book a demo to see a patrol and its audit trail in simulation.