Two platoons a year apart
Two reconnaissance platoons in the same brigade run the same platform. Same base model, same class of devices, enrolled the same week. Watch both plan a screen line along the same stretch of border and the gap is impossible to miss. The first platoon's assistant already knows which crossing floods in March and what the platoon's own SOP says about drone reports on the northern net. It drafts the reporting plan in the platoon's format with the platoon's call signs. The second platoon's assistant gives careful generic answers that could apply to any border on earth.
The platoons did not diverge on talent or budget. They diverged on one habit. For a year the first platoon has closed every mission by feeding its records and its after-action review back into a unit-scoped adapter. Its model has read every patrol the platoon ever ran. The second platoon wrote the same AARs, printed them, and filed them. Same platform. A year apart in capability.
Chatham House studied how Ukraine keeps out-innovating a larger enemy and found the answer sitting at the unit level. Frontline formations run their own feedback loops and push lessons into use within weeks, because waiting for an institution to digest the lesson means fighting last month's war. The side that learns closest to the fight learns fastest. EdgeLance builds that loop directly into the software the unit already carries.
The pallet that teaches no one
Walk into a battalion supply cage and look at the pallet by the door. Shrink-wrapped, banded, tagged for the records depot. Inside are six years of after-action reviews. The ambush that worked and the convoy that got hit. The fix a platoon sergeant improvised at 0300 and wrote down so the next man would not have to learn it the hard way. The pallet ships next week. No one will ever read it again.
That pallet holds the most valuable training data in defense. It is ground truth about how one specific formation fights on one specific piece of terrain against one specific enemy. No dataset a vendor can buy or scrape comes close, because no vendor was on that patrol. And today it trains nobody. The unit files it, the model the unit carries never sees it, and the knowledge retires with the sergeant who wrote it down.
Cloud AI makes the waste worse. Every query a unit sends to a cloud model is a training signal, and the vendor keeps it. The unit pays for inference and donates its operational patterns in the bargain. The flywheel spins. It just spins in someone else's building, and the unit that generated the data gets a subscription invoice for the privilege.
What an adapter actually is
EdgeLance distills mission records and AARs into a unit-scoped adapter. An adapter is a small fine-tuning layer, a few hundred megabytes at most, that rides on top of any open-weight base model. The base model stays untouched and general. The adapter carries the unit. Distillation runs on hardware the unit already owns, inside the same capacity policy that governs every other workload, so the flywheel turns without a cloud connection and without a data science cell.
What goes in becomes behavior. The platoon's terrain shorthand. The adversary's observed patterns in this sector. The unit's own SOPs and call signs and report formats. After distillation the model does not look these things up. It answers the way a soldier with three tours in the sector would answer. Ask about the northern crossing and it already knows the March flood, because the unit's own records taught it.
Adapters are signed and versioned in the same provenance chain that covers every other artifact in the mission application layer. A commander can read the manifest and see exactly which missions trained version 12, who approved the distillation, and which base model it targets. Adapters hot-swap onto the base model in minutes. So when a better open-weight model ships, the unit's year of learning moves onto it and keeps compounding.
An adapter you can burn
Learned behavior becomes a liability the moment a device is captured. An adapter trained on a unit's missions holds the unit's patterns, and patterns are exactly what an adversary wants from a captured device. So adapters live under the same mission burn lifecycle as everything else on the node. When a mission burns, any adapter trained on that mission's records can burn with it, and the provenance chain produces cryptographic proof that it did. The burn is an auditable event with a signature, a timestamp, and an approver.
This is what separates ownership from access. A unit that queries a cloud model can delete its account and hope. A unit that owns its adapter can produce a signed record of what trained the artifact and the exact moment it ceased to exist. That is the working definition of owning your intelligence instead of renting it, and it holds whether the reason for the burn is a compromised device, a classification change, or an order from higher.
The moat belongs to the formation that earned it
The loop closes on its own once it starts. Better missions produce better records. The records distill into better adapters, and the adapters produce better missions. The Modern War Institute argued that drone dominance comes from letting the squad fail and learn on its own terms, because a lesson centralized three echelons up arrives too late to matter. The flywheel is that argument turned into software. The squad fails, the record captures it, and the adapter remembers it on the next patrol.
And the unit owns every turn of the wheel. EdgeLance never takes the data home. Training runs on hardware the unit controls, the adapter lives in the unit's provenance chain, and the vendor could vanish tomorrow without the flywheel losing a turn. Big defense AI sells the same model to every customer, which means no customer ever pulls ahead by using it. A unit-owned adapter inverts that. The advantage accrues to the formation that did the patrols, and it accrues nowhere else.
The pallet still ships to the depot next week. But the first platoon's records now live twice, once in the binders and once in the adapter riding its base model. The second platoon saw the gap and started its own flywheel in April. Its first adapter distilled last month from eleven patrols and three AARs. Version 2 trains this week.