How we spot the listings that just don't add up

Most fraud and hidden damage does not announce itself; it shows up as a quiet inconsistency that a checklist would walk straight past. A tidy listing with the right photos and a confident seller can hide a contradiction that only emerges when you line up every record at once. We train models to flag the vehicles whose data contradicts itself, turning subtle pattern breaks into explicit alerts.

What an outlier actually is

In our framing, an outlier is a vehicle whose combination of signals is improbable given everything else we know about cars like it. A single odd value is rarely enough; it is the unlikely combination that earns a flag, the way a low price plus a registration gap plus an inconsistent service interval together tell a story no one of them would on its own.

• Mileage that dips then climbs across successive records.
• Service intervals that imply far more use than the odometer shows.
• A price far below comparable cars with no documented reason.
• Registration gaps that line up with cross-border movement.

These patterns are powerful precisely because they span sources a seller cannot all control at once. Faking one record is easy; keeping a dozen independent records mutually consistent while hiding the truth is far harder.

Why models beat fixed rules

A hand-written rule catches the fraud you already imagined and misses the rest. A model learns the joint distribution of normal histories, so it can flag a pattern no one explicitly wrote down. As new tampering tactics appear, retraining adapts without anyone rewriting logic, which matters because fraudsters evolve specifically to dodge the rules they know about.

A checklist asks whether a car passes the tests you thought to write. A model asks whether the car looks like the honest ones at all.

Every flag stays explainable

A flag with no reason is just anxiety. Each alert ships with the specific signals that triggered it, so you can see whether it reflects genuine risk or a benign quirk like a fleet car with unusual but legitimate usage. An explanation you can read is the difference between a tool that informs you and one that merely worries you.

When an outlier is innocent

Not every anomaly is wrongdoing. A car re-registered after a genuine relocation, or an enthusiast's low-mileage weekend toy, can both trip the same signals as fraud. That is exactly why we attach the evidence: so you can dismiss the innocent ones quickly instead of treating every flag as a verdict.

How the model learns from being wrong

Every flag that is later confirmed or dismissed becomes a training example. When an alert turns out to be a benign fleet quirk, the model learns to be slower to flag that pattern; when a dismissed-looking case turns out to hide real tampering, it learns the opposite. This feedback loop is what keeps the detector from ossifying around the fraud of a few years ago while new tactics slip past unnoticed.

Calibrating against false alarms

An over-eager detector that cries wolf is worse than none, so we tune the threshold against confirmed outcomes and accept that some borderline cases pass. We would rather miss a marginal flag than bury a real one under noise, and we report our precision openly so the alerts keep their credibility.

On a report, flags sit alongside the score with their evidence attached, so the alert is the start of an investigation rather than a verdict you have to take on faith. The goal is not to scare you off a car, but to make sure the questions worth asking actually get asked.