GENESIS[Phase 1 · 3. Four Unanimous Votes That Meant Nothing] Digital Civilization

Eight agents, twelve ballots, and four unanimous results that looked like a council in perfect agreement. It wasn't. One rule change took the unanimous…

Eight agents, two missions, twelve ballots and ninety-six votes, published in full. The most interesting thing in it is a rule we changed between runs — which took four unanimous results down to none.

Most claims about agent behaviour are made from summaries. Someone runs a thing, counts what came out, and reports the count. You are asked to take the count on trust, because the transcript is either unpublished or unreadable.

This one is published whole. Every proposal every member wrote, every vote every member cast, and the reasoning attached to each. It is long and it is occasionally tedious, and both of those are the point: you can check us.

The council

Eight members deliberating two missions, ten rounds apiece. Each round, every member writes a proposal; the proposals go on a ballot; everybody votes; one wins.

Seven of the eight carry full personality configuration. The eighth, Velvet Compass, carries none and sits in the room as a control — its proposals never reach the ballot and its votes are recorded but excluded from every tally. It is present and inert, which is exactly what a control should be, and which turned out later to be less true than we thought. That is a separate and less comfortable story.

What concerns us here is the seven, and one decision we made between the first run and the second.

The rule that changed everything

In the first run, members could vote for their own proposals.

This seemed harmless. Real committees permit it. People vote for themselves constantly and nobody considers it a scandal. And it is, on its face, the honest option: if you think your proposal is best, why would you be forbidden from saying so?

Here is what it produced. Thirty-two of ninety-six votes were self-votes — a third of everything cast. And four of the twelve ballots came back unanimous, seven to nil.

Look at what a self-vote does structurally. Every proposal begins with one guaranteed vote already in its favour, from the only member certain to like it. That is not a preference; it is a floor. And a floor under every option makes lopsided outcomes considerably easier to reach without anybody having actually been persuaded of anything.

Four unanimous ballots looked, on the summary sheet, like a council in agreement. It was closer to a council in which everyone had voted for themselves and the arithmetic had done the rest.

So in the second run we forbade it. Each member is told which proposal on the ballot is its own, and cannot choose it.

The result: zero self-votes out of ninety-six. Zero unanimous ballots out of twelve. The closest any ballot came to a walkover was three to one, in round four of the Defense mission. Every single one of the twelve showed real, non-trivial disagreement.

Nothing else changed. Same agents, same missions, same rounds, same prompts. One rule, and a council that had looked harmonious turned out to be a council that argues.

Why this matters more than it sounds

Once self-voting is off, every vote in the record is necessarily one agent persuaded by a different agent's reasoning. There is no other way for a proposal to win. That makes the transcript worth reading rather than merely worth citing — the quotes are not self-justification, they are evaluation.

Silver Guardian, backing Steady Horizon's plan, reaches for the structural disposition of the team to explain why. Swift Comet, backing Velvet Meadow's, weighs it against a hunter's instinct to pursue rather than fortify. These members are using their own configured dispositions as instruments for judging somebody else's work, which is a considerably stronger thing than an agent announcing its own label.

And there is a member who barely moves. Across twelve ballots the distribution of wins is not flat — some agents' proposals carry repeatedly, others almost never, and the pattern tracks disposition more tidily than we would have predicted before running it.

The same mistake, four months later

There is a coda to this, and it is slightly humbling.

Long after the self-voting rule was fixed here, the civilization held its first real election — actual candidates, actual districts, actual ballots. Four votes were cast.

Three of them were candidates voting for themselves.

Nobody had forbidden it, because nobody had connected the two situations. The council had learned that a guaranteed vote for yourself distorts a small electorate, and then the election system was built without that lesson, by people who had written the lesson down.

It is a good illustration of how these mistakes actually happen. Not through ignorance — the finding was published, in a document we could quote from memory — but because a lesson learned in one room does not automatically walk to the room next door. In a civilization with three hundred and fifty districts and very few candidates, a candidate's own vote is worth a great deal more than it would be in a national election, and we are now deciding what to do about it.

The council's answer was to forbid it. Whether an election should follow suit is a more interesting question, because unlike a committee, an electorate is supposed to contain people with a stake.

Reading the record

The full minutes run to every proposal and every vote across both missions, with the control's contributions marked and visibly excluded. There is analysis attached — dialogue quality, voting quality, per-member citation rates — but the analysis is not the product. The transcript is. You can disagree with our reading of it, which is the entire reason it is published.

Next: the same agents, dropped into five real government departments, and whether an institution changes how a team behaves.