Still in the Loop

On October 31, unveiling the Army's Next Generation Command and Control system, Brigadier General Steven Carpenter stated the season's position in one sentence ^[1].

So again, the human is making the decision, the final call, looking at everything that you need to do in order to prosecute a valid target within the Law of Armed Conflict — Steven Carpenter

Nothing in the record contradicts him. The open question is not whether a human still decides, but where he is standing when the decision arrives. The math that makes the sentence necessary: Carpenter counts more than 1,500 potential targets and 1,200 incoming threats a day — a volume, he says, no staff of analysts could read fast enough to matter ^[1]. The machines have taken the volume. The Army's $700 million Striveworks contract makes Chariot — a "central hub for decision-making" — the core software of the new command system, and its measure is concrete: work that takes a team of four to ten people an hour now runs in minutes ^[2]. The Air Force got there earlier and said it more plainly, setting out in its Experiment 3 to build an "automated command-and-control pipeline" meant to accelerate decisions and reduce the cognitive load on human operators ^[3]. The exercise's own observers found human judgment still sharpening the machine's recommendations, and fed those sharper assessments back into the algorithms — but the pipeline's stated purpose is to carry the load, not share it. By late October the machines were even handling the Army's people: the personnel system is being rebuilt as a "People Supply Chain," synchronizing enlistment, training, and assignment in real time the way a logistics pipeline moves ammunition, with the technical specialties — military intelligence, air defense — flagged first ^[4]. The Army now manages its soldiers in the same breath it manages its stockpiles. The verbs tell the rest. Across the autumn's documents, the human is consistently cast as the one who authorizes and can revoke — not as the one operating the system or approving each action. The June National Security Presidential Memorandum orders that autonomous weapons policies "respect the constitutional chain of command": the safeguard, in other words, is command authority, not a person signing off on each target ^[5]. Ukraine's new A1 center already runs the model live, with commanders responsible for "setting boundaries" while the Underdog system lets a drone finish the last stage of an attack on its own at better than 90 percent success ^[6]. Governance literature has already named the arrangement: "execution sovereignty" — institutional authority over what autonomous agents do, plus the power to withdraw it ^[7]. Authority without operation: that is where the season has left the human. The safeguard rests, first, on the person who holds the veto. Defense Secretary Pete Hegseth has spent the season describing the boundary itself as the problem.

Because we’ve written rules that are impossible, that are written for us to lose. They’re written for our guys to be in handcuffs. — Pete Hegseth

He has applied the same logic to the machines, warning that drone technology is advancing so fast the military's major risk is risk-aversion ^[8]. The handcuffs matter because they are now the whole of the human role: the man who holds the veto has spent the season arguing that the restraint is the error. The perimeter also runs on private code. In October, SpaceX throttled Ukraine's Starshield network mid-battle, and the speed restrictions disabled the drones and helicopters that depended on it, causing failed missions and civilian casualties ^[9]. Asked about it, Hegseth confirmed the terms of service leave the vendor the option to make those decisions ^[9]. A private company held the switch on an ally's war, and the Secretary's answer was that this is what the contract says. Google's Pentagon contracts deploy models in classified, air-gapped environments for any lawful government purpose, and a departing scientist, Andreas Kirsch, says the company has no right to control or veto how the Pentagon uses its AI, and limited visibility into how it is deployed ^[10]. Inside the classified wall, the oversight Google's own principles require becomes unverifiable. Then the paradox of Anthropic: the Pentagon blacklisted the company as a supply-chain risk for refusing autonomous-weapons use, and then deployed Anthropic's unreleased Mythos cybersecurity model for critical defensive operations ^[11]. The same fall that moved the human to the perimeter moved the perimeter onto private code — some of it written by a company the Pentagon says it cannot trust. None of this is the stated reason for the season's other number. The officer corps is being cut 20 percent, roughly 800 generals and admirals by January, with the on-record rationale being streamlining — removing "redundant force structure" and shifting resources from "bloated headquarters to warfighters" ^[12]. No one has claimed the machines are replacing those officers, and the record doesn't support saying so. But the cut lands in the same autumn as the relocation, on the same floor plan: fewer people at the desks, more load on the systems. Carpenter's sentence survives the audit intact. A human still makes the final call. What the autumn changed is the meaning of "final": it is now literal, the only call still assigned to a person. Authorize, revoke, set boundaries — the season's whole vocabulary for the human is the vocabulary of a perimeter. The loop is still there. The human has just been moved from its center to its edge.