GENESIS[One District of Two Hundred and Sixty-Three] Digital Civilization

For five runs this civilization published presidential turnout figures. One and eight-tenths per cent. Nine votes from four hundred and ninety-seven. Every…

For five runs this civilization published presidential turnout figures. One and eight-tenths per cent. Nine votes from four hundred and ninety-seven. Every one of those numbers measured something, and none of them measured the electorate. They measured how many districts had received a piece of paper.

The thirteenth run of OKAGI opened this week with a question no previous run had been able to ask, because every previous run was answering a different one without knowing it.

Here is the line that did it.

set_candidate_platform_statement(
    election_district_id=admission.backend_district_id,
    term_key=term_key, candidate_slug=slug, statement=statement,
)

One district. The presidency runs in two hundred and sixty-three.

What a voter was actually shown

A candidate for the presidency is registered on every district's ballot — two hundred and sixty-three separate candidate records, one per district, because the office is contested everywhere at once. That registration loops. It was written to loop, and it says so in its own comments: a candidate on 260 of 300 ballots is a real and materially different candidate from one on all 300.

The platform statement — the paragraph in which a candidate says what they stand for — did not loop. It was filed against the single district recorded on the admission.

So in one district out of two hundred and sixty-three, a voter saw a name and a platform. In the other two hundred and sixty-two, a voter saw a name.

The measurement that could not be argued with

Run 12 happened to poll two districts before anything else. The numbers came out like this.

district 242 — platform present — 78 voted of 148 asked — 53%
district 282 — platform absent — 0 voted of 67 asked — 0%

Sixty-seven consecutive agents, not one of them voting. Same run, same model, same minutes, same prompt. The only difference between those two groups of voters was whether the paper in front of them said anything.

They explained themselves, too, which is the part that ought to have been obvious years earlier. In district 242:

Tidal's platform shows practical fiscal responsibility and youth focus

In district 282:

No platform provided, cannot make informed decision

Five runs of numbers, retroactively explained

Run 11 elected a President on nine votes out of four hundred and ninety-seven. That figure was reported as a finding about the electorate — a civilization that turns out for its legislature and shrugs at its head of state.

Those nine voters were in the district where the statement happened to land.

Run 11 also produced what looked like the strongest evidence of all: on a single ballot paper, carrying three races, the same agents voted 99.7% for a councillor, 97.4% for a representative, and 1.8% for a president. Same sheet. Same moment. Same voter.

That looked like a finding about national versus local politics. It was arithmetic. A councillor contests one district and therefore always receives its filing. A president contests two hundred and sixty-three and receives one.

Runs 8 and 9 died of voters being handed bare names and refusing to choose between strangers. The fix was written then: no district is polled until every candidate on its ballot has filed a statement. It has been enforced ever since, per district — and each district it checks finds a filing present for its own local candidate, and reports, truthfully, that nothing is being held back.

The defect those two runs died of never left. It retreated into one race and stayed there for four more.

It was not only turnout

Electors in OKAGI follow districts won, not votes cast. Concentrate every presidential vote a civilization is capable of casting into the handful of districts that received a filing, and you win those districts and no others.

Run 11 seated a President on six electors of three hundred and fifty-three. Run 12 seated one on three — fewer electors than run 11, on nearly nine times the votes.

Both elections were legal. The United States Code provides for exactly this: under 3 U.S.C. §15(e)(2) the majority is taken from the electors appointed, not from those a state was entitled to appoint. The denominator shrinks to fit who came. Run 12's President needed two.

The statute did its job perfectly. It was holding up a result that should never have been produced.

The path that no country has had to walk

There is a sequence in run 12's log that deserves to be read slowly, because it is a constitutional mechanism executing unattended, at two in the morning, with nobody watching.

tick: NO ELECTORAL MAJORITY — 0 electors appointed of 353 entitled, 1 needed; goes to the House contingent election
tick: HOUSE DID NOT CHOOSE — quorate but no majority: the highest is 0 of 0 constituencies, 1 required
tick: PRESIDENT ELECTED — 3 of 3 electors appointed (353 entitled), majority 2

Three steps. No majority in the Electoral College, so the choice falls to the lower house — the Twelfth Amendment contingent election, the provision the United States has used exactly twice, in 1801 and 1825. The House convened, was quorate, and could not produce a majority either. And then the College resolved, on the electors it had.

Every clause in that sequence was implemented from the instrument rather than invented, and none of it had ever been exercised before in this civilization. It was written against a case that was not expected to arrive, and it arrived, and it held.

That is worth separating from the defect this piece is about. The machinery was correct. The elections it was conducting were not worth conducting, because the candidates were invisible to almost every voter — but given the votes it was handed, the constitutional apparatus behaved exactly as the text says it should.

It is an odd feeling to watch a system do something difficult and correct with material that should never have reached it.

Two presidents, both of them real

A detail that gets lost in the arithmetic: both elections seated an actual agent, with a term, in an office that had until then been held by a caretaker drawn by lot.

OKAGI's founding government is not elected. Every seat — 714 of them, plus 67 appointed posts — is filled by drawing lots at the moment the civilization begins, and service is compelled rather than volunteered. Refusing an assignment is an offence, adjudicated. The arrangement exists so that a new civilization has institutions on its first day instead of a year later.

It is meant to end, and the way it ends is per seat, not by decree. When an election fills a seat for the first time, that seat's caretaker stands down and that seat is never compellable again. Compulsion drains unevenly and correctly: the presidency stops being compellable the moment somebody wins it, while an uncontested district keeps its caretaker indefinitely.

In run 12, at the moment the three electors resolved, a founding president seated by lot was released, and an elected one took the office. The transition happened without supervision and without a single line of special-casing.

Three electors is not a mandate. But the mechanism that handed the office over is the same one that will hand it over when the number is three hundred.

The explanation that fitted, and was wrong

Before the right answer there was a confident wrong one, and it is worth setting out, because it was better reasoned than the truth and that is exactly what made it dangerous.

Run 12's early data looked like this. Agents holding a ballot with a single race on it voted for president at fifty-four per cent. Agents holding a ballot with several races voted for president at zero.

There was a mechanism ready to explain it, and a good one. The merged ballot — three races decided in one model call instead of three — had been adopted for throughput alone. Polling fifteen thousand agents separately for each race costs about forty-seven hours of a single GPU; merged, it costs sixteen. Nothing about that change was supposed to touch outcomes.

So the conclusion assembled itself. A national candidate, sharing a sheet with local races, loses to them. The presidency is crowded out by the councillor from your own prefecture. And an optimisation adopted for speed had been quietly altering the result it was measuring — which is a genuinely interesting kind of failure, and the sort of thing worth publishing.

It was wrong. The two groups were perfectly confounded with district. Every multi-race ballot in that sample came from district 282. Every single-race ballot came from district 242. "Merged versus single" and "282 versus 242" were the same split, and the variable that mattered — whether the presidential candidate had a platform there — had not been collected at all.

Run 11's within-paper figure had seemed to rule that out. The same agent, the same sheet, the same instant: 99.7% for one race and 1.8% for another. That controls for mood, model, prompt, district and voter. It controls for everything except the one thing.

The lesson is narrower and more useful than "check your work". The numbers were checked. What was not checked was the grouping — and a comparison can look perfectly controlled while being split on a variable nobody thought to record. A mechanism that explains the data is not evidence that the mechanism is real; it is a reason to go looking for the one that also explains it.

Fifteen thousand and eighteen

There is a second number worth knowing about this run, because it explains what the civilization looks like from outside while it is doing the least interesting part of its work.

Registering one presidential candidate means writing to two hundred and sixty-three ballots. In practice it was making exactly fifteen thousand and eighteen separate calls to the store, and runs 11, 12 and 13 each produced that identical figure.

Amended. When this piece was first published, that paragraph read that the figure was structural rather than variable — the cost of contesting an office everywhere at once. It was not. It was a defect, found later the same day, and the reasoning that called it structural was exactly the kind this piece is otherwise about.

There are 15,016 agents in this civilization. The number was the population, plus two.

The query asking which districts a presidential candidate contests took every active presidential district assignment and removed duplicates. But the model holding those assignments has a default ordering, by creation time, and the database was therefore asked for distinct pairs of district and timestamp — of which there are as many as there are agents. Nothing was removed, because nothing was duplicated.

It survived because every call returned 200. The endpoint is idempotent, so the ballots were always correct; no error was raised anywhere; no test failed, since the tests check that a candidate ends up registered, and it did. The only symptom was the fifty-three minutes — and fifty-three minutes had a reason ready, which is the sentence that kept it alive.

The fix is three characters. Fifteen thousand and eighteen becomes two hundred and sixty-three.

Those calls used to open a fresh encrypted connection every time. Each one cost about eight tenths of a second, of which six tenths was the handshake — a negotiation with a machine on the other side of the Pacific that had been negotiated with moments before. Fifteen thousand handshakes, all to the same host.

From outside, this looks exactly like a civilization that has stopped. The graphics processor sits at zero, because none of this is thinking. No rows appear in the database, because nothing is being decided. The log says nothing at all.

Run 11 spent ninety-eight minutes in that state and was diagnosed twice as a hang. It was not hung. It was queuing, politely, fifteen thousand times.

Holding one connection open for the life of the process took it from eight tenths of a second to two, and from fifteen thousand handshakes to one. Fifty minutes became thirteen. Nothing about the civilization changed; it simply stopped standing in its own queue.

That repair was real and it was also, in hindsight, a fifty-seven-fold waste made four times faster rather than removed. Both numbers improved. Neither was questioned. A reproducible number is not a correct one — it only means the same thing is happening every time.

It is a small piece of plumbing and it belongs in this story for one reason. An idle processor beside a slow system is not permission to accept the slowness. It is a measurement: whatever is taking the time, it is not thinking. That narrows the search enormously, and it is the opposite of the instinct to make the model work harder.

Why it took until now to see

The evidence had been sitting in the database the whole time, in the reasons agents gave for abstaining. It could not be read, because until the day before run 13 opened, every race on a merged ballot paper was filed under one shared reason.

An agent voting for its representative and abstaining on the presidency produced one sentence, and that sentence was stored on both rows. So a presidential abstention carried the reason the agent had given for voting in a different race — four hundred and eighty-eight refusals, each annotated with enthusiasm for something else.

Once each line kept its own reason, the first three-race ballot to arrive read like this:

house_representative | VOTED | Azure's platform aligns with my values
president | ABSTAIN | No platform statement to evaluate
house_councilor | VOTED | Azure's platform aligns with my values

One agent. One sheet. One of the three unable to say anything about its candidate, and now able to say why.

What run 13 is

A presidential candidate's platform now reaches every district it runs in. Two hundred and sixty-three filings instead of one, from a single generated statement — the paragraph belongs to the candidate, not to the district, and generating a different one per district would let a president promise different things in different places, which is a separate idea and not this one.

So run 13 asks the question the last five runs were assumed to be answering:

What is this civilization's real presidential turnout?

One district, properly served, produced fifty-three per cent. Whether that holds across a country is unknown. It may collapse. A national candidate standing in every district may genuinely mean less to a voter than the councillor from their own prefecture, and if so that is a real finding and this run will be the first entitled to state it.

The second question follows: does the Electoral College stop being degenerate? Three electors of three hundred and fifty-three is not a mandate, it is an artefact. If §15(e)(2)'s reduced denominator is doing constitutional work here rather than papering over a bug, run 13 is where that shows.

The uncomfortable part

Five runs of published presidential findings were wrong. Not imprecise — wrong, about a different thing than they claimed to be about. They were internally consistent, they were reproducible, and they were reported with confidence.

What broke it open was not a test. The entire test suite passed throughout, because the tests were written from the same understanding as the code: file a statement for an admission, check a statement was filed. Correct, and blind to the fact that one admission can mean two hundred and sixty-three ballots.

It was broken open by an agent being allowed to say, on its own line, why it would not choose.