We publish what is not working yet.
Anyone technical evaluating a risk vendor establishes the limitations in the first conversation. Publishing them costs us nothing and saves everyone a meeting.
Open limitations, as of today
- There is no validated risk score. Event weights are unfitted placeholders. Nothing we produce has been correlated with claims or collisions.
- Camera-derived signals are not shipped. Time headway, event attribution and lane discipline are designed and in build. The validated layer today is kinematic — and kinematic-only is commodity, which we say plainly on the product page.
- The validation sample is small. 53.4 km of consumer-device footage at a 53 km/h mean. It shows the pipeline computes what it claims to compute; it says nothing about a population of drivers.
- The API is not generally available. The documented interface is being implemented against the validated detector, with no uptime commitment.
- Thresholds are not universal. Our own measurements show rates moving 6.7× across defensible thresholds. Any rate we quote is meaningless without the threshold beside it.
Measurement discipline is the product.
Every number carries its conditions
A rate without a threshold, or an accuracy without a baseline, is not a measurement. Our own evaluation reports a constant-prediction baseline beside every model metric, because a small error on a mostly-straight road can mean the model learned nothing.
Derived signals get a second derivation
Anything computed from ego-motion is cross-checked against an independent route to the same quantity. A sign error in a curvature derivation does not crash — it produces confident, backwards numbers.
Tests target plausible wrongness
The suite concentrates on bugs that yield believable numbers rather than exceptions: mirrored axes, trip-boundary leakage, unit confusion between steering angle and curvature.
We would rather lose a deal than overstate
The API returns calibration: "uncalibrated" on every response until the day
it is fitted against real outcomes. That field exists so a number cannot quietly become a
score.
Where we are going
Fleets first, because a fleet owns its own collision record and can therefore prove or disprove our signal without waiting for anyone. Insurance second, entered with a model already calibrated against real outcomes rather than a pitch. We are not building components for vehicle manufacturers; that market requires a functional-safety case we do not have and are not pretending to.