Evidence

What has actually been measured.

Numbers produced by running the event layer end to end on real driving data. Not projections, and not a benchmark we selected after the fact.

Sample

Sourcecomma1M
Segments60
Frames72,001
Distance53.4 km
Sample rate20 Hz
Mean speed53 km/h

Event rates at 3.0 m/s² (0.31 g)

EventCountPer 100 km
Harsh braking1324.4
Harsh acceleration1833.7
Hard cornering2139.3

29 of 60 segments contained at least one event. Median event peak 3.61 m/s², median duration 1.42 s.

The finding that matters

Rates move 6.7× across thresholds every programme would defend.

ThresholdgBraking / 100 km Acceleration / 100 kmCornering / 100 km
2.0 m/s²0.2074.980.686.2
2.5 m/s²0.2539.352.561.8
3.0 m/s²0.3124.433.739.3
3.5 m/s²0.3620.616.924.4
4.0 m/s²0.4113.116.911.2
4.5 m/s²0.4611.211.25.6
5.0 m/s²0.5111.29.41.9

Where the acceleration mass actually sits on this sample: |a_long| has a 50th percentile of 0.27 and a 99th of 3.49 m/s²; |a_lat| has a 50th of 0.08 and a 99th of 3.03. The common 3.0 m/s² default therefore lands almost exactly on the 99th percentile of both — the steepest part of the distribution, where the count changes fastest.

Two vendors observing identical driving can legitimately report rates differing by most of an order of magnitude. This is why we publish thresholds beside every number, and why we do not publish a risk score.

Derivation integrity

The check that catches a backwards signal.

Curvature is derived from fused ego-pose, where a sign or axis error produces confident, plausible, wrong numbers. It is verified against curvature computed independently from the position track.

Trajectory vs yaw-rate correlation, this sample−0.927
Same measure, independent 551-segment run−0.917
Curvature unit plausibilitypass

Two independent samples agreeing to 0.01 is what gives confidence the derivation is not mirrored. The correlation is negative by construction — trajectory curvature is positive counter-clockwise while yaw rate is positive clockwise.

What this sample does not show

  • 53.4 km is a correctness demonstration, not a population. It shows the pipeline computes what it claims to compute.
  • comma1M is consumer-device footage at a 53 km/h mean — mixed city and motorway, averaging 0.89 km per segment.
  • Nothing here supports a claim about how often drivers in general brake hard, and nothing here has been correlated with collisions.