The vehicle has been parked since Friday and it has driven nine kilometres

The mileage is real arithmetic on unreal movement. A receiver that is standing still does not compute the same position twice; each fix lands a few metres from the last, and in a yard hemmed in by steel-clad units or under a canopy it lands much further out, because the receiver is solving with reflected signal. If your distance figure is the sum of the gaps between consecutive fixes, every one of those metres is added and none is ever taken back. Over a weekend that is kilometres, and the vehicle never left the bay.

The reason speed filtering does not rescue you is that the two fields disagree. A receiver estimates speed from the Doppler shift on the carrier, not from the change in reported position, and when it is stationary that estimate is honestly zero. So the feed says 0 km/h while the latitude and longitude keep changing, and a pipeline that gates trips on speed but sums distance on position happily accrues kilometres inside a state it has labelled 'stopped'.

One vehicle, one weekend, two numbers

1. Choose a vehicle that demonstrably did not move — a unit off the road for repair, or anything parked from Friday evening to Monday morning with a keyholder who will confirm it.
2. Export every raw position report across that window with its accuracy field, its speed field and both timestamps. Do not use a rendered trail; it has already been thinned.
3. Sum the great-circle distance between each consecutive pair. That number is what an unfiltered pipeline would have billed.
4. Compare it against the device odometer delta over the same window, if the unit reports one. Where it does not, the honest comparator is zero.
5. Repeat for a vehicle parked in the open and one parked among containers or under cover. The difference between the two is your multipath exposure, and it is usually a factor of several.

Run this before anyone argues about vendors. A pipeline that invents 9 kilometres over a weekend on one vehicle is inventing a proportional amount on every vehicle, all the time, and the error is invisible on a moving truck because it hides inside a number that was going to be large anyway.

Position error is symmetric and distance is not. Every metre of noise in a sum of gaps pushes the total up, so the mileage a fleet reports is biased high by construction, never low.

Where the stationary noise comes from, ranked

The fourth row is the one people find counter-intuitive: raising the ping rate makes phantom mileage worse, not better. Twice the fixes means twice the noise terms in the sum. That is one reason a fixed global reporting interval is the wrong instrument, and why cadence is better modelled as a function of what the vehicle is doing — the argument in ping cadence set by what the vehicle is doing.

The filter stack, in order, and what each stage costs

Order matters, because each stage assumes the one before it has run. Applying them in a different sequence produces a system that is only partly protected, and partly protected here means the leak stays open on exactly the vehicles that park badly.

Stage 2 depends on how the ignition signal was actually wired, which varies by install far more than by product. A movement-inferred substitute for real ignition sense will report the vehicle as running whenever the accelerometer twitches, and that is precisely the condition you are trying to filter — the failure modes are laid out in what an ignition event actually proves. Stage 5 is where raw distance and road distance part company for good; the mechanics of that conversion, and which of the two belongs on an invoice, are covered in map matching, turning pings into roads.

Numbers that have already been published

Fixing the filter changes history, and someone downstream has been using the old figures. Mileage-based maintenance schedules will suddenly show services as further away; driver mileage claims will drop; any distance quoted to a customer or a subcontractor will move. Recompute the affected history in one pass rather than letting old and new coexist, version the calculation, and publish the filter parameters alongside any distance figure so two systems can be compared at all. The same discipline applies wherever a computed number leaves your boundary and lands in someone else's report, which is the failure pattern in the carrier says it sent the status and nothing moved.

Be careful with anything automated on top of the number. Rules that raise a maintenance job, close a timesheet or approve a mileage claim without a human in the path will act on the inflated figure silently and repeatedly — the reason we scope this kind of agent and automation work with the data-quality gate first and the automation second. Position history is also personal data about named drivers, which is why fleets that care about where it is processed reach the same conclusion as teams weighing running language models on their own infrastructure.

Making it an install test, not an incident

• Add an overnight drift test to install acceptance. A newly fitted unit that accrues distance while the vehicle is locked up has failed, whatever the wiring diagram says.
• Test the awkward parking, not the depot. The unit that behaves in an open yard is not evidence about the one in an underground bay.
• Alert on the pattern, not the incident. Any vehicle accruing distance with ignition off for more than a few minutes should raise a data-quality flag, because that condition is never legitimate outside a tow.
• Recheck after firmware changes. Reporting behaviour while parked is exactly the thing a vendor tunes without telling anybody.

Two failure modes look like this one and are not. If a vehicle accrues no distance because it has stopped reporting entirely, you have a stale-device problem instead, handled in the eleven vehicles that have not moved since morning. And if drift is concentrated on one device model or one install batch, treat it as a rollout defect rather than a physics problem — the staged approach with data-quality exit criteria in fitting trackers across a mixed-age fleet exists to catch exactly that. Both sit alongside this page in telematics and the live vehicle data plane, part of our logistics and mobility work.

• telematics
• data quality
• mileage
• fleet operations