Who Holds the Machine When Light Is Too Slow

Start with a fact nobody can repeal. A signal from a low-earth-orbit sensor to a ground station and back, routed through a command node and a human, takes time — and in a fight measured in seconds, that round trip is the difference between intercept and impact. Against a hypersonic track or a saturating drone swarm, the honest engineering answer is that a human in the loop on every decision is not fast enough. That is not a preference. It is the speed of light and the geometry of the battlefield.

So the architecture is pushed toward autonomy at the edge: let the machine decide inside the window a human can't act in, and keep the human on the loop rather than in it. Everyone technical already knows this. What almost no one is saying out loud is the consequence — once you concede that the machine must act faster than a person can intervene, you have moved the real decision from the operator to whoever wrote the rules the machine follows, and to whoever can physically reach the machine to stop it. The kill switch didn't disappear. It moved.

Read the brief plainly: Meridian is tasked to study the battlefields of the future and the capabilities the force must master, from under the earth to beyond the moon, across AI, autonomy, directed energy, robotics, and biotechnology. Pair that with a standing four-star command whose job is to scale autonomous systems, and you have the two halves of a single decision — one group deciding what the machine should be able to do, and one chain of command owning what it does every day.

The membership is the tell. You do not put the leaders of a launch-and-compute empire and an autonomous-weapons company on a future-of-warfare panel to study incrementally. You do it because the mesh they are describing — proliferated satellites, edge compute, autonomous shooters, a software-defined command layer — is the one their companies are already building. That is not a criticism; it is the realistic way frontier capability gets into a force. But it sharpens the governance question rather than softening it: the people best equipped to build the autonomous mesh are, structurally, the people least positioned to be its independent check.

The reassuring phrase in every briefing is that a human stays 'on the loop' — supervising, able to intervene, able to abort. I want to be precise about what that phrase does and does not buy you. On-the-loop control is real only to the extent that three things are true at once: the human can see what the system is about to do, in time to matter; the human has an action that actually stops it; and the communications link carrying that action survives contact with the enemy.

In a degraded fight — jammed links, spoofed sensors, a satellite layer under attack — all three of those can fail at the exact moment they are needed. If the link is down, 'on the loop' quietly becomes 'the machine is on its own, operating on the last instruction it was given.' That is the scenario the governance has to be written for, and it is the scenario briefings skate past. The civilian hand on the kill switch is only as real as the worst-case link that connects it to the machine.

There is a quieter dependency underneath all of this, and it is timing. An autonomous mesh only behaves like one machine if every node agrees on what time it is to the nanosecond — that shared clock is how sensors fuse, how shooters deconflict, and how a human's abort command is sequenced correctly against what the machine is already doing. Today that timing comes overwhelmingly from GPS, which is exactly the signal an adversary jams or spoofs first. This is why assured, GPS-independent timing is becoming a command-and-control problem rather than a utility. A growing field of alternative position, navigation, and timing (PNT) work is aimed squarely at it — quantum time transfer, which uses entangled photons to distribute precise time without depending on a vulnerable GNSS signal, is one example (the startup Xairos is pursuing it) — because the mesh needs a clock that survives when GPS does not. If light is too slow to ask permission, the least you can demand is that the whole mesh still agrees on the clock when the enemy takes GPS away. Lose the timing and 'on the loop' stops meaning anything, because the operator and the machine are no longer even reasoning about the same instant.

I keep having a version of this conversation with a major general who lives in this problem, and the thing we come back to is not the technology — it is accountability. His point, which I think is exactly right, is that autonomy does not remove human responsibility; it relocates it earlier, into the design, the rules of engagement encoded in software, and the test regime that decides what 'the machine behaved correctly' even means. The uniform still owns the outcome. What changes is where in time the decision that produced the outcome was actually made.

That reframing is the whole ballgame for governance. If responsibility moves upstream into code and doctrine, then the oversight has to move upstream too — to the acquisition decision, the model that was fielded, the rules it was given, and the audit trail that lets someone reconstruct, after the fact, why the machine did what it did. You cannot supervise at machine speed in the moment. You can only govern the machine before it is handed the authority to act, and account for it afterward. Everything real about civilian control now happens on either side of the engagement, not during it.

Here is where this connects to the larger argument I have been making. Software diffuses; the machine does not. An autonomous-warfare mesh is not a clever algorithm — it is satellites, launch capacity, edge compute, secure interconnect, and the industrial base to sustain all of it. Whoever owns that physical layer owns the real authority, because they are the ones who can field it, update it, and — critically — reach in and stop it. The kill switch is not a line of code. It is a supply chain and a command relationship.

That is why consolidating the autonomous mesh inside a handful of private platforms is a governance event, not just a procurement one. When the compute, the satellites, and the autonomous shooters increasingly run through the same few companies, the practical ability to halt the machine concentrates with them too — unless the government deliberately architects independent control, independent audit, and an off-switch it alone holds. Sovereignty over an autonomous force is sovereignty over the machine, physically. Anything less is a civilian hand hovering over a switch that someone else wired.

If I had a seat at that table, I would argue the hardest problem is not which capabilities to master — the panel will get that right. It is writing the control architecture before the capability is fielded, not after. That means a government-held, link-independent abort authority designed for the degraded case; assured, GPS-independent timing — resilient alternative PNT so the mesh stays coherent and the abort stays sequenced when GNSS is denied; a mandatory audit trail on every autonomous engagement so responsibility can be reconstructed; and a rule that no single private platform owns the compute, the satellites, and the shooters for the same mission thread without an independent check. Build the governance into the mesh, or you will be retrofitting it after the first incident.

The uncomfortable truth underneath all of it is that we have already conceded the first premise: in the fights that matter most, the machine will act faster than a person can. Once you accept that, the only questions left are who wrote its rules, who can physically stop it, and who answers for it when it is wrong. Those are civilian-control questions, and they are being decided right now — in a 120-day study, in a new command, and in a supply chain — whether or not we choose to call them that. Light is too slow to ask permission in the moment. So we had better decide, in advance and on the record, who holds the machine.