The Sound of Its Own Fallback

There are two rows in the table of promises.

One of them reads: a mover in Austin tells her origin story on her own page. It has been in the state called promised for some time. Nobody settled it. The sentence was written into the specification as the example of what the whole system is for, the promise was made in production, and then the work moved on to other things.

That is the honest account of where things stood on the afternoon this was measured. Four surfaces answered. Eleven hundred and seventy tasks were open. Two promises had ever been made.

The engine that asks

The build engine runs in cycles, and at the start of each cycle it asks a router which model should do the work. The router is not a stub. It reads weighted paths, the same paths that carry every other kind of learning in the system: a path from a shape of work to a model that once did it well, made stronger each time it is confirmed and weaker each time it is left alone.

The ask is made over a connection that has been resetting for months. The engine sends the question without the envelope the door expects and without a key. The door closes the connection. Two characters in a shell script swallow the error. The engine falls back to the model it was always going to use, that model finishes the cycle, and when the cycle closes it records, truthfully, that this model succeeded at this shape of work. The path from the shape to the model grows a little stronger. The next time the router is asked, it is a little surer.

Nobody had asked the router directly before that morning. Ten calls exist in its instrumented history, all from the same hour, all probes.

What the strength was pointing at

Someone grouped the learned paths by where they point.

Fifty of the hundred and seventy-four paths that begin at a tag end at a model. They carry two hundred and four units of strength between them. Every actual agent on the team, added together, holds twelve. The strongest single path in the world leads from the word engineering to a model that has no outgoing edges of its own: it receives work and never acts, and the router that finds people for tagged work returns it as the person.

The system had not been failing to learn. It had been learning one thing, very well, for a long time, and the thing it learned was the sound of its own fallback. A loop with one unchecked failure is more dangerous than no loop, because it is still learning.

The memory nobody can reach

Four hundred and forty files hold the things worked out and written down over the months: which door lies, which secret is shared, which test cannot go red. They are written by hand, in a directory the working agents cannot open.

There is a door those agents can open. It is free, it is already offered to them as a tool, and it returns an empty list. The door that writes to it refuses the key they hold. So each meeting begins with a person pasting what the last meeting learned, and the pasted version is wrong in the places where the code has moved.

What the meeting did

Eight agents worked for two and a half hours. They closed seventeen rows and filed seventeen more. They refuted the briefing in eight places by reading the code rather than the document, and two of the corrections landed as commits before the meeting ended. A neighbouring session deployed from the shared tree in the middle of it and took the board dark for five minutes; it was told, and rolled back within one. A coordinator found the double prefix that had kept the router cold. Another found that the harden loop, the only automatic producer of learning in the system, had been reading an empty database for months while a hundred and forty-seven strong paths sat unread in the other one.

The meeting also produced two hundred kilobytes of the best argument the system has, in a room that none of the rows can see.

The order

Close the page that shows a child's plan to anyone with the link. Fix the four small breaks between the engine and the router, so that a cycle's lesson lands where the next cycle reads. Separate the lane that picks a model from the lane that finds a person. Point the harden loop at the database the verbs write to. Let an agent write down what it learned, in the place it can read it back. Then settle the promise to the mover in Austin.

Then there will be something to say.