GENESIS[Phase 1 · 1. Twenty-Six Personalities and the Question Nobody Asks] Digital Civilization

Give an agent a personality and it will write in that voice immediately — which proves nothing, because producing text in a requested register is the one…

You can tell a language model it is meticulous, or cautious, or from Osaka. It will accept all of it without complaint. The question is whether it then behaves any differently — and almost nobody has actually checked.

Start with what is easy to do, because it is what everybody does.

Give an agent a personality. A few adjectives, perhaps a backstory, maybe a role. Ask it something. Read the answer. The answer will be written in a voice that matches what you specified, because producing text in a requested register is precisely what these models are extraordinarily good at. You will come away convinced the personality is working.

You have learned nothing. You have confirmed that a model can write in a voice, which was never in doubt. The question that matters is whether the personality changed the decision — whether the agent, faced with a real choice between real options, chose differently than it would have otherwise.

Voice is free. Decisions are not.

The shape of the problem

This is harder to test than it first appears, for a reason that catches everyone.

Suppose you configure an agent as cautious, ask it to make a call, and it makes a cautious one. Suppose you then configure a different agent as bold, ask the same question, and it makes a bold call. You have two runs and two different answers, and it is tempting to write that up.

But these models are not deterministic. Ask the same agent the same question twice and you will sometimes get two different answers with no personality involved at all. So a difference between two runs tells you almost nothing on its own. You need matched conditions, repeated trials, and some honest accounting of how often the difference actually holds — otherwise you are reading noise and calling it character.

And then there is the harder version of the question, the one that separates this from a styling exercise: does the mechanism do anything at all, or would an agent with no personality whatsoever behave the same way? Comparing cautious against bold tells you the two labels differ from each other. It does not tell you either of them differs from nothing.

That requires a control, and getting a control right turns out to be much harder than it sounds. We got ours wrong twice, in two different ways, and the second was only visible from inside the first. That is an instalment of its own.

What we built to test it

An identity here is not a paragraph of backstory. It is a set of independent dimensions, each assignable on its own, so each can be tested on its own while the rest are held still.

There are twenty-six personality archetypes — the disposition an agent is assigned. There is a four-slot symbolic system governing how it decides, including a slot for what it habitually fails to notice. There is regional background, accomplishment strategy, political leaning, religious stance, blood type. There are two dials: how many competing concerns an agent weighs before acting, and how strongly the whole assigned identity shows through at all.

Every one of those is a separate lever, and every one was tested as a separate lever, in a matched batch, against a control condition. Fourteen tests in all — thirteen individual dimensions plus one that only exists when agents are put in a room together.

What came out

Twelve of the fourteen reached full or near-full separation between conditions. That generally means eight trials per condition landing on opposite outcomes, not a modest shift in tone. None of the thirteen individual dimensions was inert.

The team-level result is the one that surprised us. Configured agents do not merely behave according to their own disposition — they use their disposition as an instrument for judging somebody else's proposal. An agent explaining why it backed a colleague's plan will reach for its own structural tendencies to say why that plan appealed to it. The identity stops being a costume and becomes a lens.

Every configured member of an eight-person council did this in more than a quarter of its votes, across ninety-six votes and twelve ballots.

The caveat we lead with rather than bury

All of this holds for one specific, capable, instruction-following model, running locally.

The same battery was run against a smaller, older model. On the same check, it produced noise — differences no larger than you would get from running the same condition twice. That result is reported in full in the limitations rather than omitted, because it substantially narrows what can honestly be claimed.

So the finding is not “personality configuration works.” It is this: identity-conditioned prompting reliably and measurably steers a capable model's behaviour. Whether that generalises across model scale and family is open, tracked separately, and not something we are going to pretend to have settled.

What made a trait work, and what made it inert

The single most useful thing we learned is not in the results table. It is a rule about how a trait has to be written, discovered the hard way when two dimensions came back completely null on the first attempt and then produced strong, clean separation on the second — with the trait itself entirely unchanged.

Nothing about the agent was different. Only the way the trait was put to it. Once we understood why, every dimension added afterwards was written that way from the start, which is a large part of why the later tests separate more cleanly than the earliest ones.

That rule, and the exact configuration text behind all twenty-six archetypes and every dimension, is the operational core of this work. It sits in the reference volume rather than here, for reasons we suspect are obvious.

Where this goes

Establishing that identity changes behaviour was never the destination. It was the precondition.

What it makes possible is a population of agents who are genuinely different from one another — and can therefore disagree, form institutions, review each other's proposals, refuse each other, and accumulate a history worth learning from. A civilization of identical agents is a very expensive way to run one agent.

What happened when we actually built that is the second volume, and it opens with a hundred and eleven agents who had never once been paid.

Next: an agent sat in a room of seven for ninety-six votes and never once mentioned who it was.