About $1.80M to Cohere Solutions — evaluating the strategies behind what is trained on, taking data rather than the model as the subject

1The subject of evaluation is on the data side

What the description names is not the performance of a model but the strategy behind the data used to train it: what to gather, how to select it, in what order to learn from it.

The same model behaves differently on different material. Whether a difference in performance comes from the design of the model or from how data was chosen is hard to separate from outside. Evaluating the data side independently reads as taking on that separation.

2Space command and control as the domain

The domain named is space command and control. Unlike land or sea, means of observation are limited and the objects move under physical law, while situations change quickly.

In such a domain, material obtained from actual events is scarce. The question of what to train on can itself set the ceiling on performance. Placing data strategy at the centre of evaluation reads as built on that constraint. The record does not describe specific methods or criteria.

3What the record supports

What can be read is that the Air Force awarded it, who the recipient is, that obligations are about $1,799,804, that the period runs two years from 2026-04-07 to 2028-04-07, and that the subject of evaluation is the data strategy in training.

Across the 222 federal spending records this site holds as of 2026-09-01, outlays are zero on 155, positive on 66 and negative on one. This record sits at zero, and little time has passed since it began. Gauging progress from the outlay figure is not possible at this stage.

4The model side, or the data side

Evaluating the same AI, the question changes with which side is taken as the object. This contract takes the second.

The same model behaves differently on different training material. Whether a difference in performance comes from the design of the model or from the choice of data is hard to separate from outside. Space command and control, named as the domain, yields little material from actual events, so what is chosen as training material can itself set the limit on performance.