One estate, four terrains

A national transportation number averages over places that do not resemble each other. This is the finding of TR-2026-34 made operable: pick a terrain class and see the quantity that actually decides automation there, and which of the constraints is in the way. Every figure is from the review.

What binds, across all eleven hypotheses

The review stated eleven hypotheses before assembling evidence, and each resolved to a binding constraint. H6 resolves on two, so twelve counts cover eleven hypotheses. The distribution is the most portable result in the paper.

Nothing on this estate binds on AI algorithms. One hypothesis binds on computation efficiency, and it sits downstream of a compliance rule rather than upstream of a capability. Whether a machine can do the work is settled more often than whether the work can be bought, measured or reached.

The test that could have overturned all of this

The review included a hypothesis capable of destroying its own organising choice: if the automation-relevant quantities inside a terrain class sort by ownership rather than by slope, terrain is the wrong partition. The test quantity is the share of bridges in poor condition across four steep-terrain states, against the national 6.68 percent.

Two of the four sit below the national figure, so the partition survives only partly, and the paper says so. One state carries the class.

Source: Physical AI and the Department of Transportation (TR-2026-34). Constraint vocabulary: the eight.