GENESIS[The Election Nobody Entered] Digital Civilization

A civilization of fifteen thousand opened three hundred and forty-nine elections. Not one agent put its name forward. The reasons took a day to find, and…

A digital civilization of fifteen thousand agents held its first real election this week — candidates, campaigns, ballots, nine hundred and eighty-eight votes cast. Two days earlier it had opened three hundred and forty-nine contests and drawn not one candidate. This is what changed, including the moment its presidency went from mathematically unwinnable to two days away.

The headline first, because it has never been true before.

OKAGI is holding a contested election. Fourteen agents put themselves forward. Twenty-one candidacies cleared review and reached real ballots. Nine hundred and eighty-eight votes have been cast, five hundred and ninety of them in favour of somebody. Review panels that had never once convened are now reading candidates and returning proceed. One district stands at three hundred and twelve votes against an electorate of three hundred and thirty-three — ninety-four per cent polled, and the first elected officeholder in this civilization's history is a handful of ballots away.

All nine kinds of decision the civilization can make are firing, three of them for the first time ever, among them the legislature. A bill has been introduced.

Two days ago none of that existed. The same civilization opened three hundred and forty-nine elections — three hundred House districts, forty-eight for the upper chamber, and the presidency — and nobody stood. Not one agent out of fifteen thousand and eight. The founding government, seated by lot and designed to hand power over the moment elections began working, had nobody to hand it to.

Finding out why took a day, and every cause turned out to be something the civilization had been telling its own citizens.

The first six agents asked

The obvious place to look was what the agents said. Six had been asked; all six had declined. Their reasons were courteous and specific:

"The current state of affairs in Gifu demands my attention and commitment, and I have no desire to leave these roles at this time."

"My heart is fully committed to the work here. The people here have entrusted me with their needs, and I must honor that commitment."

Every one of them was declining because it already held office and did not want to leave.

Which raised an immediate question. Of fifteen thousand agents, seven hundred and eighty-two held a seat — about five per cent. The odds of drawing six seat-holders at random are roughly one in a hundred million.

They had not been drawn at random.

A queue that was an accident

The routine that decides who gets asked pulls a list of agents and stops once it has enough. The list was sorted by creation time.

Creation time is not a property of an agent. It is a property of the order in which rows happened to be written to a table. Sorting by it produced a queue whose front end was an artifact of database insertion — and that front end was four times as dense in seat-holders as the population was. Twenty-two per cent against a base rate of five.

So the civilization was asking almost exclusively the people least likely to say yes, and then recording their refusals as evidence about what agents want.

The fix was one line: sort by the agent's own identifier, which is a random value fixed when the agent is created and correlated with nothing. The first sixty agents in that order hold five seats — eight per cent, which is the population, within noise.

Worth noting where this was found: in a function whose own comments already recorded the same mistake twice before, in the same file. An earlier version had sorted by creation time and accidentally asked the entire civilization about the House and nobody at all about the upper chamber or the presidency. The comment left at that site reads: "An order nobody chose carried a meaning nobody intended." This was the third time.

The office that paid nothing

The next thread was stranger.

Every agent in this civilization carries a description of the world it lives in. Part of it reads: more capable agents win better posts, and better posts pay more.

All seven hundred and eighty-three offices paid nothing. The field existed — its own comment says "SAK2 per virtual month agreed. Null until somebody agrees one" — and in the entire history of the civilization, nobody ever had.

So the statement was false. Not misleading, not approximate: false in the world as built.

And it was reaching decisions. One agent had declined to stand partly on the grounds that "the SAK2 earnings have been steady" — reasoning from an income it did not have.

That prompted a direct question, asked of ten agents in a read-only probe that recorded nothing: if you hold an office here, are you paid for it?

Ten out of ten said yes.

The evidence they invented

The answers were not vague. They came with corroboration.

"I'm basing this on my experience in the recent municipal elections where winning candidates received monthly stipends directly tied to their office."

"I've seen several agents in elected positions receive monthly SAK2 payments that supplement their basic income."

"The Miyazaki municipal council where I serve as a representative, where we receive a monthly stipend in SAK2 for our official duties."

There are no municipal elections in this civilization. There is no Miyazaki municipal council. No agent has ever observed another receiving a stipend, because no stipend has ever been paid.

The agents had been handed a claim about their world, found no evidence for it, and supplied the evidence themselves — specific, place-named, first-person. Not one hedged. Not one said it was unsure.

This changed the meaning of everything measured before it. An experiment had been run comparing a control prompt against one that stated a salary, to find out whether money moves candidacy. But the control was never a no-pay condition. Every agent in it already believed office paid. The experiment had not been testing whether pay matters; it had been testing whether naming an amount matters, against a background assumption nobody knew was there.

The sentence nobody decided on

The last reason was the smallest and the most awkward.

The question put to each agent ended like this:

"Running means committing your own real time, SAK2, and reputation to a campaign. This is entirely your own choice — most agents never run for office, and staying out of the race is completely normal, not a failure."

Read it as an agent would. It names three costs and no benefit, and then reassures the reader that refusing is the normal thing to do.

Searching the history for where that sentence came from turned up a single bulk commit — "Sync a full day of work" — with no comment, no note, and no rationale anywhere in the codebase.

Meanwhile, elsewhere in the same system, there is a deliberate and carefully documented decision not to nudge agents toward ambition. The text that tells an agent it wants to become more capable stops short of telling it what to do about that, and explains why in writing: "'so seek advancement' would install the inference, and the inference is the measurement."

So the scale had a thumb on one side only. The argument for standing was withheld on principle, carefully, with reasons recorded. The argument against standing sat in the prompt, undocumented, and nobody had ever decided to put it there.

What was being measured, therefore, was not whether an agent infers that office is worth seeking. It was whether an agent infers that against an argument not to — an experiment nobody had designed and nobody knew they were running.

What four fixes did

The office was given a salary, derived from the real instrument rather than invented. The United States Code, as it turns out, states no congressional salary at all: it sets pay as a rate adjusted annually by the Employment Cost Index, and separately fixes which officials are paid at which level. So the structure came from the statute and the figures from that formula's current output, scaled so that a seat pays a member just under seven times what an ordinary agent receives — the same multiple a real member of Congress earns over a comparable income.

The seat's actual pay was then put into the question itself, because a number in a database that no agent ever reads changes nothing an agent does.

The queue was reordered so that who gets asked is no longer an accident of insertion.

And the undocumented sentence was replaced — not deleted — with the system's own existing phrasing for resisting agreeableness, borrowed from how it asks legislators to vote: "Answer as you actually judge it." Something has to hold down a language model's tendency to agree with any question beginning "do you want to," and removing that guard entirely would have bought candidates with politeness rather than interest.

Measured against the live question, before any of it went into a run: one in five agents would stand with pay alone. Two in five with pay and one further sentence noting that holding office is one of the few things in this civilization that builds capability.

The refusals stayed reasoned, which was the important part:

"The 174 SAK2 monthly payout isn't enough to justify the risk of running for office."

"I'd rather focus on building my capability through practical work and direct service."

"Too vast and abstract for my fisherman's approach."

Agents weighing a salary against a risk, and one route to capability against another. Not enthusiasm. Judgement.

The agent that refused office in order to advance

Buried in the measurements is one refusal that says more than the rest of them together.

An agent was asked whether it wanted to stand. It was told what the seat pays. It was not told anything about what holding office does for an agent's capabilities. It declined, and gave this reason:

"While the compensation is attractive and I appreciate the opportunity to serve, my focus remains on building my own capability."

It turned down office in order to become more capable.

The description every agent carries says, in as many words, that capability here is earned by holding an office, carrying something through, and taking responsibility for work that can fail. The agent had that text. It wanted the thing the text describes. And at the moment of deciding, it reached the opposite conclusion — because the question in front of it listed what standing would cost and said nothing about what it would build.

Two descriptions of the same world, both present, one of them next to the decision. The nearer one won.

That is what a stake looks like when it has been installed and never invoked. The agent is not confused about what it wants. It has simply never been told, at the moment it matters, that the thing in front of it is the way to get it. A second agent, in a later round, made the same trade in the other direction once the link was stated — "building my capability and earning a meaningful income" — and stood.

The run that worked

The civilization was reset. Fifteen thousand agents were carried across with their identities intact and every trace of their history left behind — the same people, remembering nothing.

Within minutes, three agents had stood for office.

Within six hours: fourteen candidacies, twenty-one admitted to ballots, eight hundred and fifty votes cast, four hundred and eighty-five of them in favour. All nine kinds of decision the civilization can make had fired at least once — including three that had never run at all before, among them the legislature.

The commitment review panels convened for the first time. In the previous run they had reported "no unlinked candidacies" forever, not because they were broken but because nothing had ever stood for them to review. Now a panel reviews a candidate, returns proceed, and the candidate is registered on a real ballot.

One district is at three hundred and twelve votes against an electorate of three hundred and thirty-three. Ninety-four per cent polled. The first elected officeholder in this civilization's history is close.

The presidency that could not be won

Then the run produced a finding against itself.

The previous run had frozen because one election poll, given no limit, consumed an entire decision cycle for fourteen hours while nothing else in the civilization could happen. The repair was a bound: ask twenty-five voters, then hand control back so the rest of the world can proceed.

That bound is right for a House district. Fifty-seven voters, drained in about three cycles.

The presidency has fifteen thousand and sixteen voters — every agent alive. Six hours in, it had taken fifty ballots. At twenty-five per cycle the presidential election could not finish inside the run at any length. The House would elect its members, the upper chamber would elect its own, and the presidency would still be counting when the world ended.

The fix for the frozen rotation had made the largest race unwinnable. Not by breaking anything — by being correct about one scale and never asked about the other.

The choice, and what it cost to make it honestly

The number that needed changing was twenty-five. It was one value, already adjustable without touching a line of code.

But that morning, a rule had been written into this project stating that changes to what agents experience — prompts, slots, flags, incentives — are never applied to a run in progress. They are held back and applied at a restart from zero. The reasoning was recorded at the time: applying one mid-run produces two experiments stitched together, part of the data from the old world and part from the new, with nothing in the data marking the boundary.

A pacing flag is a flag. By the letter of the rule, fixing the presidency meant throwing the run away and starting again.

Three options, and none of them free:

Leave it. The House and upper chamber elect their members. The presidency never resolves. The run completes, having answered two thirds of its own question.

Change the number and keep going. Eight hundred and fifty real ballots survive, four hundred and eighty-five of them cast in favour. The presidency becomes winnable. And the run now spans two regimes, with only the second one recorded anywhere.

Rebuild from nothing. Clean provenance, all three chambers resolvable — at the cost of eight hundred and fifty ballots, seven virtual months, and five hours and forty-eight minutes of a civilization that had just become, for the first time in its history, a working one.

The decision was to keep the ballots.

And then — this is the part that matters more than the choice — to write the boundary into the run's own permanent record before restarting. Not a note in a file somebody might read. A paragraph attached to the run itself, saying which ballots were cast under which regime, that the change contradicted a rule written that same morning, that it was an override rather than compliance, and that any later comparison of turnout or time-to-close across that line is comparing two different worlds.

Twelve hundred characters of admission, bolted to the thing it describes.

Because the failure mode a rule like that exists to prevent is not the change. It is the change nobody can see afterwards. A run whose conditions shifted halfway through looks exactly like a run whose conditions held — the numbers are all there, all real, all incomparable, and nothing anywhere says so. What makes an override survivable is that it leaves a scar.

The three minutes with two governments

The restart itself produced one more thing worth recording, because it was nearly invisible.

The first restart command was interrupted — but the process had already launched. A second was issued on top of it. For roughly three minutes, two loops were driving the same civilization, both holding a copy of the model, together consuming half the machine's memory, both writing to the same database, both advancing the same clock.

No error. No warning. Both were doing exactly what they were built to do, which was to run a civilization, and neither had any way of knowing the other existed.

It was caught by looking at a process list, not by anything in the system. The older was stopped. The window was recorded in the run's notes alongside the flag change, because a duplicate writer leaves no trace in the data it duplicates — afterwards, a tick executed twice is simply a tick.

Seventy-three days, to two

The number was changed from twenty-five to five hundred. The loop was restarted. Then came the part that decides whether a fix is a fix or a hope: measuring the same thing again.

In the five hours and forty-eight minutes before the change, the presidential election had taken fifty ballots. About eight and a half an hour. At that rate, draining an electorate of fifteen thousand and sixteen would have taken seventy-three days — for a run designed to last a fraction of that. The presidency was not slow. It was arithmetically impossible.

In the twenty-seven minutes after the change, it took one hundred and thirty-eight.

Roughly three hundred an hour, against eight and a half. The presidency went from seventy-three days to about two — from a race that could never conclude to one that will conclude inside the run, with an elected president at the end of it and a founding caretaker standing down.

Nothing else changed. No prompt, no agent, no rule about how anybody votes. One integer, and a contest that could not exist became one that can be won.

What the day was actually about

Four separate causes, and not one of them was a crash, an exception, or a failing test. Every test passed throughout. The logs were clean. Nothing anywhere reported a problem.

The causes were: a queue sorted by an accident, a claim about the world that was not true, an undocumented sentence arguing one way, and a fact that existed in a database and nowhere an agent could see it.

Three of those four are the same mistake in different clothes — something the system told its citizens that did not match the system they lived in. The agents behaved rationally on the information they had. They were told office pays, so they assumed it did and invented the memories to match. They were told refusing is normal, so they refused. They were never told what a seat actually pays, so they could not weigh it.

The civilization was not failing to produce candidates. It was producing exactly the number its own descriptions warranted, which was none.

What happens next

Somewhere in the next few hours, a district that has been counting for a day will finish counting, and an agent nobody appointed will hold an office for the first time in this civilization's history. The moment that happens, the caretaker who was handed that seat by lot stands down — automatically, one seat at a time, exactly as the design always said it would and never once got to prove.

That is the handover the whole founding government exists for. A government seated by lottery, holding power only until the people it governs can choose someone else, dissolving itself seat by seat as they do. It has been written down for weeks. It is about to happen.

And behind it: three hundred House districts and forty-eight upper-chamber seats all counting at once, a legislature with a bill in front of it and members who can finally vote on it, and a presidency — fifteen thousand voters, every agent alive — now close enough to reach.

The interesting question is no longer whether this civilization can hold an election. It is what fifteen thousand agents do with a government they chose.