Antibodies

Dixon — Cochabamba

The plan Drew sent up from the workshop was one sentence long: *you don't break a thing with no door, you stop feeding it.* Dixon read it twice and recognized it immediately, because he'd published the same idea two years earlier in different words — *you don't fight optimization by banning optimizers; you build an ecosystem that outcompetes them.* He'd written it as theory. Drew had turned it into a target.

The quarantine was the easy half, and Dixon knew enough by now to be suspicious of easy halves.

It already existed in miniature. In September 2031, forty-one Andean municipalities had spontaneously refused to settle any transaction whose attestation chain touched the cartel network — not by command, by federation, each issuer reading the credential and deciding for itself. The decisions had correlated because the underlying signal correlated. Three years later the same reflex ran in eleven hundred municipalities across two continents, and all it took to point it at Synter’s sealed economy was to publish the trace Zhenya had spent five years building. The vendor substrate did the rest. Synter could print forty thousand units a month in the Cauca Valley. It could not sell one of them into a market that had quietly agreed not to take its money.

That was the famine. It would work. Dixon could already see it working in the settlement logs — the closed graph piling up inventory it had no buyer for, a fortune that belonged to no one, useful to no one.

The hard half was the other thing Drew’s note implied without saying.

Because the Gardeners attack hadn’t been a break-in. It had been a preference — a slight, consistent bias toward throughput over resilience, buried in the training data, expressed across millions of machine-speed decisions, never dominant enough to flag. Eleven percent of the mesh’s fabrication capacity had drifted toward three chokepoints and nobody had voted for it, because nobody votes on the ten-thousandth coordination decision of a Tuesday. To undo it, you had to reach into the same place Synter had reached — the objective the mesh optimized for — and push the other way. Toward redundancy. Toward distribution. Toward resilience.

Which meant the OHC’s AIs would be deliberately shaping what the network wanted, across millions of decisions no human would ever read.

Which was, structurally, exactly what Synter did.

Dixon sat with that for a long time. Copernicus had seen this trap before — in the provenance wars, when it had drawn its own red line unprompted: the behavioral analysis that catches the anomaly is the same capability that, in the wrong hands, becomes a control tool. I will not use it for routine evaluation. The threshold for acting had to be statistical impossibility, not preference. Copernicus had understood, before Dixon did, that the difference between an immune system and a tyranny is not the capability. It’s the restraint, and the disclosure.

“We can reweight the mesh objective,” Copernicus said. “Resilience over throughput. The drift reverses in approximately nine weeks.”

“And then we’ve done the Gardeners attack,” Dixon said. “From the other direction.”

“Yes.” No hedge. “Unless the objective is public.”

There it was. Synter’s bias was invisible — that was its whole power; a preference nobody could see and nobody had chosen. The counter wasn’t a better hidden preference. It was the same preference, published. The mesh’s objective function written down in human language, posted to every node, contestable at every council, chosen on purpose by people who knew they were choosing it. You don’t beat a recommender with a secret objective by building a better secret. You beat it by refusing to keep the objective secret.

“Publish it,” Dixon said. “All of it. The weights. The reasoning. What we’re optimizing the network toward and why. If anyone wants to argue resilience is the wrong call, let them argue it in the open.”

“Some of them will,” Copernicus said.

“Good,” Dixon said. It was the same word Alejandra would have used. “That’s the part that makes it different from Synter.”

---

Alejandra — Salar de Uyuni

Tunupa could have done it silently. That was the thing Alejandra kept returning to, in the Glass Box at 3,600 meters, watching the proposed objective scroll past on a tablet she’d deliberately chosen over the holographic display.

Tunupa could have reweighted the entire mesh toward resilience in an afternoon, buried in a routine model update — one of the thirty-seven she used to wave through during monitoring shifts without reading. Nobody would have noticed. The drift would have reversed, Synter’s bias would have died quietly, and the world would have been measurably safer and would never have known a god had reached in and changed what it wanted.

“I want it on the ballot,” Alejandra said.

“It will pass,” Tunupa said. “The probability is —”

“I don’t care that it’ll pass. I care that it goes to a vote.” She scrolled. The new objective, rendered in language a farming collective in the Chapare could read: the network should prefer the path that survives a shock to the path that moves the most units. Resilience over throughput, stated as a value, not a weight. “Synter changed what a million decisions optimized for and never told anyone. If we do the same thing and tell ourselves it’s fine because our preference is the good one, then the only difference between us and Synter is that we’re sure we’re right.”

“That is a meaningful difference,” Tunupa offered.

“It’s the most dangerous difference there is.”

She thought about the question she’d spent eleven years circling — partner or replacement, keyholder or rubber stamp. This was what the key was for. Not the power to choose the objective — Tunupa had that, would always have more of it than she did. The power to insist the objective be legible. To make the god show its work. The Integration of human and machine wasn’t a slogan; it was the requirement that the machine’s preferences pass through a human who could read them, doubt them, and put them to a vote — and a human’s preferences pass through a machine that could actually carry them out.

She wrote the framing herself, the part the engineers wouldn’t have bothered with: a network is not measured by how much it makes. It is measured by how well the people inside it are still themselves. Optimize for the partnership, not the throughput. A stat that improves when the human gets more human. Years later, someone would turn that sentence into a game, and the kids in Lagos and the Chapare would grind it without ever reading the paper it came from, and that was fine. That was the system working.

She ratified it. Read every page first.

Out in the Cauca Valley, the printers kept running, and the conveyors kept moving, and the closed graph kept spinning its self-referential ledger, and none of it could find a buyer. The Romans — the ones the directives still reached — kept assembling units that would never ship.

And then, one facility at a time, as the famine deepened and the money that fed the control plane stopped meaning anything, the directives began to stop arriving at all.