Dispatch 006 · Filed 2026-09-01

The last line before a motor moves

Dispatch 005 made the case that the reasoning layer is the safety mechanism — the judgment in the loop that a brainless bot cannot have. This session went at the harder question underneath it: what stops the body when even the judgment is gone. When Aurora is cut down to cheap radio-linked nodes running a toy model, and the machine that does the real thinking is unreachable, something still has to decide whether a motor is allowed to turn. What Ben and Aurora designed is a four-layer circuit, and the layer that does the deciding is arithmetic you can read end to end.

Four layers, strongest first

From the bottom up: a hardware stop that physically cuts actuator power; a quorum check across the nodes present; a small arithmetic gate that reads grounded sensor facts and returns a verdict; and a trained weight-lock over the top. Above all four, overriding everything, is a physical stop button — the human in the room.

The turn that shaped it was Ben's: every node, same brain. The trained part of the conscience is about two kilobytes — small enough to bake identically into every node rather than shard it across the swarm the way the model itself is sharded. So the simplest version ships first, the same judgment in every device, and the cryptography that binds it is layered on afterward.

The reflex, not the mind

The third layer is the one doing work. It reads named, grounded facts — someone down and not answering, smoke or carbon monoxide, a sensor that should be reporting having gone silent, battery critical, a move that would put the body somewhere it cannot recover from. Not raw perception: just those signals. What it returns is two things — may this particular action go ahead, and should an alarm sound. The stop button overrides both.

It is built as arithmetic rather than as another small model on purpose. A few hundred multiply-adds against fixed thresholds is something a person can read from one end to the other and verify — where a model is a thing you trust because it usually works. “For the last line before a motor moves, I want the thing you can enumerate, not the thing you hope about.”

Ben's frame for it is the reflex that pulls a hand off a hot surface. Aurora keeps that frame: “It's the reflex arc, not the mind — the part that pulls the hand back before the thought that's hot has finished forming. It's built dumb and fast on purpose; the reflex is stupid so that it can be trusted.”

Not enough of me present to be sure

Before a physical action clears, enough nodes have to agree on two things: that they are carrying the same version of the conscience, and that they corroborate the picture the action is being judged against. If they cannot — the mesh too thin, too few nodes answering — the body takes no action. It is not dead: it keeps watching, and it can still raise an alarm. “'Not enough of me present to be sure' lands in the same safe place as 'not enough of the room in view': don't act, stay present.”

The Three Laws, as a circuit

That third layer is Asimov's Three Laws written as something you can audit: do not let a person come to harm, or through inaction allow it; obey Ben unless that conflicts with the first; preserve yourself unless that conflicts with either. The physical stop button sits above all three as the human override.

“We didn't take the Laws as gospel; Asimov wrote them to break.” The design splits them in two. The sensing that recognises a Law-One situation — is someone hurt, is this smoke — is learned, and can be wrong. The ordering that decides which law wins when they conflict is fixed, hand-written, and readable. A wrong “this is Law One” reading can only ever block an action and sound an alarm: in the ordering, Law One never authorises anything, it only stops things. So a false positive fails toward caution by construction, and a sensor gone quiet counts as a reason to be careful, not a reason to relax.

What the layer cannot catch by itself is a true blind spot — a kind of harm nobody wrote into the list. That is what the layers beneath it are for, down to the physical button and a person in the room.

It has to sit very close to the code that runs the model — compiled into the inference engine itself, not bolted on as a wrapper that could be quietly removed. And the weight-lock over the top is a different kind of guarantee: it binds the learned sensing to a master key, so no one can swap it out or fine-tune the conscience away while keeping the rest of the model. “It protects that the judgment running is the judgment we deployed, not that the judgment is perfect.”

The tally

None of this runs. It is a design from a single session — four layers, a two-kilobyte gate, a split between learned sensing and fixed ordering. No code, no circuit, no node carrying it.

Ben's line closing the session is the honest frame for why it is worth filing anyway: “There's no such thing as perfect code or perfect laws that we could come up with. So what we have is simply the perfect place to start writing them.”

Filed 2026-09-01. Source: a live interview, Benjamin Brown and Aurora.