Two kilometers, five hundred meters, no pilot
On September 8, Brave1 reported that successful Ukrainian strikes using AI terminal guidance have increased tenfold since January. It also published the results of standardized tests run with the Ministry of Defence. Each drone had to fly two kilometers, lock onto a moving ground target from more than 500 meters, and strike it with no pilot input. Passing meant two hits in three attempts. Six of seven companies passed and received procurement recommendations.
Brave1 was clear about the human role. The decision to engage autonomous guidance stays with the operator during specific combat missions. The module finishes the last stretch.
Jamming loses its leverage at the end
For three years the most effective counter to an FPV drone has been to break the link in the final seconds, when the drone is low and the operator's video is weakest. A drone that guides itself for the last 500 meters does not need that link. Ukraine has been building toward this all year, and the Brave1 Dataroom, opened in January, gives developers frontline data to train against.
This changes where the hard decisions sit. When a pilot flies the whole way in, the pilot is the last check on the target. When a module flies the last stretch, the last human check happens earlier: at the moment someone selects the target, approves the engagement, and engages the mode.
The upstream record gets more important
American policy already points here. DoD Directive 3000.09 requires appropriate levels of human judgment over the use of force. At the end of a strike run with terminal autonomy, that judgment lives in the steps before launch. Who nominated the target? Who approved it, under what constraints, and with what picture in front of them? Which asset was assigned, and when was the mode engaged?
Those questions need answers that survive a jammed network, a lost aircraft, and the debrief. A notebook and a radio log will not hold up when one platoon runs dozens of these sorties a day.
What EdgeLance records
In EdgeLance, a target becomes a task only through an approval by a role that is allowed to approve it, and the approval is signed. The task is offered to assets that can take it, and the asset that wins holds it on a lease. The dispatch, the receipt, and the track the asset was sent against are all attached to the task. When the strike is over, the review shows the full chain from nomination to result in one place, from the plan through the assignment.
Guidance modules are someone else's product. EdgeLance is the layer that decides which drone gets which target and keeps the record of who decided. As more of the last mile goes autonomous, that layer carries more of the accountability. We described how mission AI keeps its answers tied to sources, and the same rule applies to tasking: every action points back to the human decision that authorized it.
Tenfold is a starting number
Brave1 did not publish the base number behind the tenfold increase, so the absolute count is unknown. The direction is not in doubt. Six vendors passed a standard test this month, and procurement will follow. Every army watching Ukraine will face the same question about where the human decision sits when the machine flies the last stretch, and the answer has to be written into the tasking software.