We checked our own design three ways. Ten problems came out.
A tracker for trucks, designed in two nights
On 3 and 4 November we designed a product of our own: a tracker for heavy trucks and vans, Morocco first. It reads the vehicle on its diagnostics connector, decides on board with small edge models, and sends what matters to a datacenter.
No truck has been tested. Every decision below is still a proposal that a person has to accept or reject. We publish the design journal anyway, because of what happened between version 1 and version 2.
Research changed the product before any drawing
Research came first: 38 sources and 7 written constraints. Three facts reshaped the design.
Our search found no commercial LTE-M or NB-IoT offer in Morocco. The low-power radio that most trackers rely on may not be available here, so the design uses LTE Cat 1 bis until an operator tells us otherwise.
As we read CNDP deliberation 17-2014 on vehicle geolocation, a driver's maximum speed may not be processed. Driver scoring therefore scores harsh events and idling. Not speed.
A cab reaches 70 °C in the sun. Lithium cells charge only up to about 45 °C.
Version 1 looked serious
A lithium backup cell for eight hours of reporting after a power cut. One SIM operator. Firmware updates installed when the vehicle is parked. A review by people who agree with each other would have passed it.
We checked it three ways instead. A state exploration walks through every situation the device can be in. A numeric check tests the budgets: current, capacity, temperature. A resilience analysis removes one component at a time and watches what falls. Ten findings came out.
Three checks pointed at the same weak spot from three sides. The device alone could not promise the power-cut alarm.
Version 2
The guarantee was split in two. A supercapacitor, rated for the heat, sends one last alarm. The datacenter declares a tracker silent after missed heartbeats.
Proving that design produced a new counterexample. Unplug the tracker and plug it back: the ignition messages counted as heartbeats while the alarm was still waiting. The rule that came out of it is now part of the protocol. Every message carries the pending critical events first, and a heartbeat counts only once it has reached the network.
The rest followed. Two firmware slots, with installation only when parked on vehicle power. Reserved slots for critical events, with 48 hours of queue proven to fit in half the flash. A sleep current target of 2 to 6 mA at 12 V. Two operator profiles on an eSIM.
Thirteen claims now hold, each with a certificate. Four properties were also checked against a deliberate defect, and each one caught it. A proof that cannot fail proves nothing.
From proof to firmware
The proven state machine was generated into C++17. Five truck situations replayed on a workstation pass 13 checks. The same core, cross-compiled for a small microcontroller board, is a 125 KB image for the pilot rig.
One lesson from that step. A sensor reading enters the model's world at the next tick, so "one step after the ignition comes on" in the proof is two ticks of real time. One test in the truck plan measures exactly that.
What is not proven
A proof covers the model, not the truck. The sleep current is a target and nobody has measured it on a bench. The radio choice rests on a search, not on an operator's answer. And two single points of failure are accepted for the pilot: the device itself, and the one datacenter site.
What remains is physical. Build the rig, measure the sleep current, run the tests in a truck.