How much of this can you trust?

Every derived metric on this site, with its formula and a signed account of what is known to be wrong with it. Nothing here is a “data may be incomplete” disclaimer — each figure states which direction its known problems push it.

Derived metrics

A metric is here when a numerator meets a denominator, because that is where a number can be wrong in a way you cannot see. Each carries its formula, a signed ledger of what is known to be wrong with it, and the query behind it.

In registry order, headline metric first, because this is not a ranking. Sort by name to read it as a lookup table instead. Each metric links to its own page.

Crash reports per million driverless miles	reports / million miles	Per mile driven, how often does this operator file a crash report?
Severity mix (5-Day share)	% of filings	Do these operators report crashes the same way?
ADS and ADAS are different datasets	reports	Why can't these two crash numbers be added together?
Authorized automated vehicles	vehicles	How many robotaxis is each company legally permitted to operate?
No rows match this filter.

The nine marks

Uncertainty is drawn, not disclaimed. These appear identically everywhere and are never explained again in prose.

Ordered as the marks are numbered in the code, so the reference reads the same way every time.
Mark
	A period still accruing filings. It will rise.
0	Observed, and genuinely zero. Not the same fact as no data.
	Null. Never a zero, never interpolated.
	Self-reported by the operator, no external check.
	How much of the record a rate can speak for.
driverless_deployment	Which population an exposure figure counts.
	A numerator with no divisor. Never reads as safe.
2026-06	Month resolution only — the source has no day field.
D1	How checkable a divisor is: D1 regulator-filed, D2 self-published, D3 proxy, D4 absent.
No rows match this filter.