In 2025, road deaths in Great Britain fell by 4 per cent while the number of people killed or seriously injured rose by 4 per cent — the first year on record in which the two headline measures of the same phenomenon moved in opposite directions. This report argues that the insurance market has been reading outcome data as though it were a scoreboard, when standardised loss statistics record only failures and cannot observe the losses that effective controls prevented; a flat or diverging line is therefore a net position, not evidence of unchanged risk. Testing three plausible upward pressures against primary data, we find the speed-pressure mechanism unsupported — extreme speeding contracted roughly three times faster than average speeding between 2018 and 2024 — while distraction survives on roads and cannot be tested elsewhere for want of any instrument. Reviewing what historically changed outcomes, one property recurs: the largest gains came from controls that reduced reliance on human performance at the moment of exposure. But two investigated losses, the Ocado Andover fire and the Tempe automated-driving fatality, show that in each case the relevant control existed, functioned, and was deliberately suppressed — meaning an underwriter asking whether a control is present receives an accurate answer that conceals the operative fact. From this we set out a testable proposition: an operating unit can lose the capacity to detect and arrest abnormal conditions without producing more initiating events, and where that occurs the first observable insurance consequence may be conditional severity rather than claim frequency. It has not been demonstrated in claims data, and we specify the study that would settle it — but the practical conclusion does not depend on it, because moving underwriting from whether a control exists to whether it is available, activated, performing, overridden, degrading and independently evidenced improves risk information either way.