GENESIS[Phase 1 · 4. Nought of Seven, Then Seven of Seven] Digital Civilization

The same seven agents, five government departments, consecutive missions. On the Commerce ballot not one vote invoked the agent's own personality. On the…

The same seven agents were dropped into five real government departments and told to deliberate. Forty votes were cast, thirty-five counted, and the two most interesting ballots sat next to each other: one where nobody mentioned who they were, and the next where everybody did.

Establishing that a personality changes what an agent does is one thing. It is a laboratory result — a matched batch, a controlled comparison, a clean number. The harder question is whether any of it survives contact with an institution.

Put a team inside a government department, give it a real brief with real constraints, and the department may simply overwhelm the individuals. That would be a perfectly respectable outcome, and arguably the one you would want from a civil service. It would also mean that everything established about personality stops mattering the moment agents are put to work.

So: the same seven agents, five institutions, one mission each, flat peer vote. Election Administration first, then the Executive Office, then Commerce, then the Treasury, then Artificial Intelligence.

The scoreboard

Five ballots, and no two alike.

Steady Horizon carried Election Administration four votes to seven. Silver Guardian took the Executive Office six of seven — a strong consensus, level with the Treasury as the firmest of the five. Commerce was the mess: Clear Raven won it with three, in a genuine three-two-two split with a margin of one. The Treasury went to Velvet Meadow, six of seven. And Clear Raven took Artificial Intelligence with four.

Forty votes were cast in total and thirty-five counted — the difference being the control member, whose votes are recorded and then excluded, and whose proposals never reach the ballot at all.

Of the seven regular members, three never authored a winning proposal in any of the five institutions. Clear Raven took two. The distribution is not flat, and it does not look like chance.

The two ballots that sat next to each other

Every vote was scored for whether the agent's stated reasoning explicitly invoked a personality dimension — its disposition, its symbolic decision style, its regional background. Across the whole run, eighteen of thirty-five counted votes did.

The distribution of those eighteen is the finding.

At Commerce: nought of seven. At the Treasury, the very next mission: seven of seven.

Same agents. Same configuration. Consecutive ballots. On one, not a single vote was scored as reaching for an identity; on the next, every single one was.

We want to be careful about what that means, because the temptation is to overclaim it. The zero is a scoring outcome, not a vow of silence — agents at Commerce still wrote in their own registers, and at least one vote there does mention a communication style in passing. What the pair shows is that identity surfaces in stated reasoning very unevenly, and that the thing which moves it is not the agent but the room.

A commercial brief and a treasury brief pull different things out of the same seven people. That is a more interesting result than a flat rate would have been.

Where the agents actually looked

The most revealing detail in the whole run was not scored at all, and we nearly missed it.

Silver Guardian, proposing at the Treasury, wanted to brief “the Governor's office.” Clear Raven, at Commerce, proposed preparing “a detailed report for Congress.” Elsewhere in the run the agents reach for Homeland Security and for a Ministry of Finance.

USPE has legislative bodies. It has governor chambers. What it does not have is Congress, Homeland Security, or a Ministry of Finance — and the agents were not reaching for USPE's institutions at all. They were reaching for the ones in their training data.

Asked to behave institutionally, an agent will furnish the institutional apparatus it already knows, and it will do so confidently, in the middle of an otherwise sensible proposal. Nothing in the brief invited a Ministry of Finance. It simply arrived, because that is what a serious government memo contains.

This is worth more than the vote counts. It says something about what “acting institutionally” means to a model asked to do it: the shape comes from somewhere, and if your civilization does not supply the shape insistently enough, it will be supplied from elsewhere.

The control, and what it did not do

Velvet Compass sat in all five, with no personality configuration at all, its votes recorded and excluded.

Across the five institutions it agreed with the team's winning proposal exactly once, at the Executive Office. Four times out of five it backed something the room did not choose.

At the time we read this as the control being visibly out of step — configured agents converging on each other while an unconfigured one drifted. We would put it more carefully now. That control was later found to have been built wrong in two separate ways, and the comparison it supports is narrower than we believed when we ran it. The vote counts stand; the interpretation is the subject of a considerably less comfortable instalment.

A bug that flattered us

One more thing came out of this run, and it arrived as an embarrassment.

The metric that decided whether a ballot counted as unanimous was checking whether the winning proposal's votes equalled the total votes cast. With the control's vote recorded but excluded from the tally, the totals never matched, and ballots that were in fact unanimous among the members who could vote were scored as merely strong.

A one-line comparison, a ceiling nobody had noticed, and a set of figures that was quietly wrong in our favour in one direction and against us in another. We found it, fixed it, and re-derived the numbers above.

It belongs in the article because it is the kind of thing that never appears in a results table. A metric that produces plausible numbers is very difficult to distrust, and the only reason this one came under suspicion was that a human looked at a ballot, counted the votes by hand, and got a different answer.

What it settles

Personality survives institutional context. It does not survive it uniformly — the same agents produce wildly different rates of identity-driven reasoning depending on which department they are standing in, and the two extremes of that range sat on consecutive ballots.

What an institution does not do is flatten them. Three members never won anything across five departments; one won twice. Whatever the room does to these agents, it does not make them interchangeable.

Next: the same five institutions, re-run with a department head who could overrule the vote entirely. Five chances. Zero overrides.