What a driverless fleet actually costs per occupied mile, who loses which jobs at which penetration, and the legal categories that quietly stop applying. 168 of 168 audit checks pass.
Working Notes · calibrated case study, revision 3 · PDF · 24 pages · 1.2 MB
A bottom-up cost, labour and policy model of autonomous ride-hail in the Denver-Boulder region: human benchmark built from wages and the IRS mileage convention, a six-component AV cost stack across three conditional technology states, a queueing sub-model for remote supervision, separate short-run demand and long-run ownership channels the model refuses to sum, and a VMT-neutrality threshold the modelled fleet never meets. Every parameter carries a source tier and a range; the audit harness runs 168 checks.
[ figures ]
The evidence, drawn.
fig. 01 · price per occupied mile, human vs machine
Epochs are conditional technology states, not forecast dates. The human line is built bottom-up from wages and the IRS mileage convention and is the model's single largest stated uncertainty; AV retail includes a 20% markup over operating cost.
fig. 02 · positive through 50% share, then the cliff
The two panels share a series and differ in scale by ×100 · printed on each frame, because hiding it would be the lie. Partial automation reads as a net local jobs gain right up until it doesn't, which makes 'it's been fine so far' a dangerous inference.
fig. 03 · the loudest line item is not the largest decline
Absolute $/occupied-mile decline, 2026→2035 states. Supervision gets the narrative attention (it is the only component with its own queueing sub-model); boring vehicle hardware is the bigger number. The model's own words: opacity is not the same as dominance.