Crashla

Explanatory Note (AI-generated text)

This page aims to compare "miles per incident" across Tesla, Waymo, and Zoox within the NHTSA time window (June 15, 2025 through January 15, 2026).

Top chart: solid line = MPI from all incidents; dashed line = MPI from nonstationary incidents; dot-dash line = MPI from at-fault-weighted incidents; error bars use vmt_min and vmt_max.
Three company charts: VMT line (with error bars) and incident bars by speed bucket, where darker sections indicate higher or unknown speed.
Tesla mileage assumptions are anchored to tracker sources (robotaxitracker.com and robotaxi-safety-tracker.com) and then aligned to this same NHTSA window for apples-to-apples comparison.

The colored band around each MPI line is a 95% Bayesian credible interval. Model: incidents ~ Poisson(λ · m), where λ is the rate (incidents per mile) and m is VMT. Jeffreys prior: λ ~ Gamma(0.5, 0) (improper). Posterior after observing k incidents in m miles: λ | k, m ~ Gamma(k + 0.5, m). MPI = 1/λ; quantiles are inverted via a monotone decreasing transformation.
The credible interval combines uncertainty from incident counts (Gamma-Poisson) and from VMT (vmt_min/vmt_max) conservatively: the lower MPI bound uses vmt_min with the upper λ quantile, and the upper MPI bound uses vmt_max with the lower λ quantile. This yields the widest possible band.
For partial months (June and January), VMT is pro-rated by the coverage fraction so that VMT and incident counts cover the same observation window.
The point estimate shown in the line is the Bayesian posterior median of 1/λ, not the simple ratio m/k. For small k (especially Tesla), the prior pulls the estimate slightly downward; for large k (Waymo), the difference is negligible.
Fault-weighted incidents (thin line): each incident contributes its fault fraction (weighted average of Claude, Codex, Gemini) instead of one full count. The sum of fractions is treated as a pseudo-Poisson count; this is a heuristic but reasonable approximation.