ORBIS 4 introduces on-board inference: running a model on a partner satellite payload or a FIELDKIT node rather than after the data has come to ground.
The motivation is bandwidth. A hyperspectral cube is enormous and most of it is uninteresting. Running a screening model on the payload and downlinking only the interesting subset, plus a compressed summary of the rest, changes the economics of high-revisit monitoring substantially.
The constraint is that you can only do this for questions you already know how to ask, using a model you are confident in, because you are throwing away data you cannot get back. We have been conservative: on-board inference currently runs screening models with high recall and mediocre precision. It decides what to keep, never what the answer is.
That distinction is the whole design. A screening model that occasionally keeps something boring costs bandwidth. A model that occasionally discards something important costs a customer their event.
ORBIS 4 also brings the unified tasking fabric across GRIDSCAN, TIDEWATCH and AGRISENSE. Release notes are in the documentation.