What the AI Does, and What It Doesn’t
A farmer says “move the bales from the top field to the yard.” The model’s job ends almost immediately: it identifies a bale, and it says which part of it the machine should take.
Everything after that is measurement. Tyina finds it against a fitted ground plane, measures its real dimensions in depth, confirms it is within what the machine can carry, checks the ground beneath it, works out the approach from its own geometry, and verifies it is carrying what it thinks it is carrying before it moves.
The model contributes a word and a rough region. The machine contributes every number.
That division is the safety case, not a preference. In testing, the model’s region was 55mm out — and the result was still correct to the millimetre, because nothing it said was ever used as a measurement. It was used to choose; the sensors did the arithmetic.
Take the network away and the machine loses the name and the description. It does not lose the ability to measure, to verify, or to refuse.
