Pre-combat checks never covered the AI
Before a patrol leaves the wire, someone checks the batteries, the radios, the fills, the ammunition, and the water. Army doctrine is specific about pre-combat checks and inspections because the cost of discovering a dead battery at the objective is measured in people.
AI is now equipment. Teams carry models for detection, transcription, translation, and answering questions about the order. Nobody checks them before step-off. Which model is on which device, which version, whether it fits in memory alongside the map and the camera, whether the documents it needs are indexed. Those questions get answered in the field, at the worst possible moment.
How unstaged AI fails
The failures are mundane. A model that fit on the test device runs out of memory on the phone that also has the map, the camera, and the radio app open, and it crashes halfway through a transcription. Two operators carry different versions of the same model, so they get different answers to the same question and nobody knows why. The question-answering tool works fine but cannot find Annex B because nobody loaded it.
These are logistics failures, and the military already knows how to solve those. You plan the load, you check it before you leave, and you write down what went out the door.
What a mission pack holds
EdgeLance stages AI the way a unit stages any other equipment, through a mission pack. Each pack is built for a class of device: a sensor node, a phone, a laptop, a forward server. It declares the models the device will run and their size, the documents and data the AI may search, the roles that can use it, and the memory budget it has to live inside.
Before step-off, a fit gate checks each device against its pack. Does the model fit with everything else this device has to run? Is the retrieval index complete? Is the pack signed by the people allowed to approve it? A device that fails the check gets flagged while there is still time to fix it, in the motor pool, not at the objective.
Same pack, same answer
Every device of the same class gets the same signed pack. Two operators asking the same question of the same documents get the same answer from the same model. When the after-action review asks why the AI recommended something, the mission record shows which pack was loaded, which model answered, and which sources it used.
Packs are model-agnostic. The unit picks the models, local or API, from whatever its program has approved, and swaps them between missions without touching the rest of the system. The pack is the contract between what the program approved and what the device carries.
The pack also carries the retrieval sources, so the mission AI answers with citations from the documents the unit approved and nothing else. If the order changes, the pack is rebuilt and re-signed, and the fit gate runs again on every device before the change goes out. Routing work between devices follows the rules in how the inference gateway works.
Put it on the checklist
The recommendation for any unit fielding AI, with EdgeLance or anything else, is simple. Add the model loadout to the pre-combat checklist. Know what is on each device, prove it fits, and record what went out. The units that do this will trust their AI because they know exactly what it is.
It also changes what a burn means. When the mission ends, the burn destroys the keys for everything the pack produced, and the record of what was loaded survives for the review. The load plan and the destruction record end up as two ends of the same document.