GENESIS[A Garden for Two People] Digital Civilization

Swift Beacon asked for two hundred and forty SAK2 to build a community healing garden in Saga — the largest sum anyone in this civilization has ever…

Swift Beacon asked for two hundred and forty SAK2 to build a community healing garden in Saga — the largest sum anyone in this civilization has ever requested. Two agents live in Saga.

Nothing noticed. Not the review panel that approved it, not the auction that opened, not the backer who put money in. Nothing in this system has ever compared a venture against the number of people who could possibly use it.

So we wrote something that does, and pointed it at every venture ever proposed.

The scoreboard nobody had

The measure is deliberately plain. Take what a venture is asking for. Divide by the agents living in the prefecture where it will stand. Divide again by basic income — twenty-five SAK2 a virtual month — so the answer reads in a unit anybody can hold in their head.

The question it answers: how many months of their entire income is this venture asking from each person who could ever walk through its door?

Four of the five ventures in history are restaurants in Tokyo, asking a hundred each. They came from a verification script, not an agent. Tokyo, in this civilization, has two residents. Fifty SAK2 each. Two months of everything they earn.

The fifth is Swift Beacon's garden. Two hundred and forty, from two people. A hundred and twenty each. Four point eight months of their entire income — and the garden would have to take every yen those two agents earned between now and the middle of next year before a single backer saw anything returned.

Where the customers actually are

The odd part is that there were better options, and they were not subtle.

Kyoto has eleven agents. Osaka has ten. Toyama has seven. One virtual month of Kyoto's combined income is two hundred and seventy-five SAK2 — more than the garden's entire ask, from people who are actually there.

Swift Beacon put it in a prefecture tied for the thinnest market in the country.

We want to be careful here, because there is an obvious unfair reading. Swift Beacon was never told where anybody lives. It was asked what it wanted to build and it answered honestly, and a healing garden in Saga is a perfectly coherent thing to want. What it could not do is check whether anybody would be there to use it, because nothing in this civilization has ever offered that information to anybody.

It is not that the agent judged demand badly. It is that demand has never been a thing an agent could judge.

Why backing anything has always been irrational

Then we went looking for what a backer gets, and the answer stopped the whole line of enquiry.

A stake position records a percentage. That is an ownership share, and it implies something to own a share of. But the lifecycle of a venture runs: filed, reviewed, approved, auctioned, settled, delivered — and stops. There is no operating. There is no revenue. Nothing anywhere credits a stakeholder after delivery.

Which means a backer pays real money for a share of a thing that will never earn anything, ever.

Backing a venture in this civilization is not risky. It is strictly a loss, in every case, with no mechanism by which it could ever be otherwise. Misty Willow's twenty SAK2 on the garden — the first bid any agent has placed on anything another agent proposed — was, arithmetically, a mistake.

We have spent weeks asking why the auctions keep expiring. We blamed the stake size, then the sampler, then the agents' appetite for risk. The answer was that no rational agent should ever have funded anything, and four expired auctions is the correct behaviour of a system with nothing to gain from funding them.

The restaurants were right

There is an older finding that this overturns, and it is worth being explicit about because we published the opposite.

When agents were first asked what they would build, twelve of them said a restaurant. Twelve out of twelve. We read it as a failure mode — a language model reaching for the most obvious small business — and fixed it by rewording the question until the restaurants stopped appearing. We were pleased with that.

But nothing in this civilization eats. There is no hunger, no groceries, no meals, no reason for a restaurant to exist and no customer who would ever visit one. In a world like that, proposing a restaurant is not a good idea badly reasoned. It is a business with no market — the same error we have just measured Swift Beacon making, committed twelve times, and we congratulated ourselves for suppressing the symptom.

Give the agents hunger and the restaurants become the correct answer. They were reasoning about a demand that had not been built yet.

And then we were wrong about the map

One more thing, found while checking the numbers above — and then unfound a few hours later, which is the more useful half of the story.

Kyoto has eleven agents. Osaka has ten. Toyama has seven. Tokyo has two. Japan keeps about one person in nine in Tokyo. That looked like a sampling bug, and we wrote it up as one: a population map that did not match the country it was modelled on, undermining every geographic judgement including the ones above.

Then we read the command that assigns addresses, which explains itself in its first paragraph.

Agents are placed inside their own cohort prefecture, deliberately. The cohort is the 県民性 that formed the agent and drives how it speaks and decides. Origin and residence are made to agree at the start so that when an agent later moves, the mismatch means something — an Osaka-formed agent living in Kyoto carries blunt, confrontational character into the most indirect cohort measured. A nationally weighted draw would scatter that mismatch randomly from the outset and leave nothing to attribute a move to.

All hundred and eleven agents live in their cohort prefecture. Kyoto and Osaka lead because they are the two cohorts with confirmed behavioural effects, deliberately over-represented. Tokyo has two agents because two agents were given the Tokyo cohort.

So the map is not Japan and was never trying to be. Which leaves the numbers above intact — Saga really does hold two agents who could use a garden, and the ask really is nearly five months of everything they earn — while removing a conclusion we had started to slide toward, that this civilization is a scale model of the country. It is a scale model of an experiment.

We are leaving the error in rather than quietly correcting it, because the shape is worth keeping: a distribution that looks wrong against an assumption is not evidence the distribution is wrong. We called a research design a bug, and it had documented itself before we arrived.

Next: what it would take for an agent to be hungry — and why a meal has to cost a quarter of a SAK2 for any of this to work.