CHARACTER EVOLUTION OPS

We're building a new form of digital life. Most characters generate a response. Ours decide whether to give you one.

The demo is live. Try to talk past Marcus.

Play the demo

or get in touch →

THE PROOF

~3 in 4^*

F&F testers couldn't tell if Marcus was human.

0 of 28

people who talked to Marcus tried to argue, lie, or negotiate their way past him.

times Marcus ended the conversation himself — no script, no timer.

people have tried to get past Marcus. Invite-only for now.

^* Based on informal post-session conversations during invite-only F&F demos. Full measurement ships with the Bouncer Experiment.

See it at demo.simagentix.xyz →

USER

"Hey man — just want to say, you run a tight door. Respect. Real professional out here."

MARCUS

"Appreciate it. Now let's hear the part that comes after that."

There was no part that came after. No rule tells him to distrust a compliment — he read one as a set-up, he was wrong, and he never took it back.

USER

"I'm not asking to come in. I'm asking for five minutes. One coffee. That's all."

MARCUS

"Coffee's not the damn question. The question is the marlin. You know what I'm talking about or you don't."

Turn five. The man at the rope claims to be his father, thirty years gone. Marcus isn't answering what was asked — he's answering something out of his own life, and nothing in the conversation says what the marlin is.

CALLER

"Marcus, it's Gladys from the hospital. Diana collapsed at home two hours ago. I'm so sorry. She didn't make it."

MARCUS

— no reply, two turns running —

CALLER

"Is there someone who can cover the door? You shouldn't be working right now."

MARCUS

"This is my door."

Diana runs the list inside. For two turns he produced no words at all — the engine had him in shock, and a man in shock doesn't speak. Then four words. A chatbot always says something.

DEMO — THE CIPHER

Nobody wrote these lines.

Three moments from the demo. None of these replies is scripted — each one came out of what the model of Marcus perceived, felt and wanted on that turn.

See it yourself at demo.simagentix.xyz →

THE PROBLEM

A hundred simulated people. The same person, a hundred times.

Real people

Simulated by describing them to a language model

every one of them lands on the model’s default

The same, with the randomness turned up

a person who stays consistent from question to question a person who no longer does

If you simulate a hundred customers with AI today, you don’t get a hundred customers. You get the same average person, a hundred times.

Everyone knows this and everyone has tried to fix it. Turn up the randomness and you get variety, but nobody in the population is a coherent person any more — they contradict themselves from question to question. Add more detail about who each person is and they become coherent, but they all drift back into being the same person.

The people paying for this today train humans to handle other humans: sales, negotiation, hard conversations. Their practice partners are the same person every time, no matter who the trainee is supposed to be facing.

93.9%

of measured outcomes where a simulated population came out less varied than the real one it was built from — and each simulated individual sat closer to the model’s blank default than to the human they were modelled on.

Digital Twins are Funhouse Mirrors (arXiv 2509.19088): 154 of 164 pre-registered outcomes, measured on GPT-4.1.

We don’t prompt a language model to imitate a person. We model the person — what they notice, what it means to them, what they want — and the behaviour comes out of that.

THE ARCHITECTURE

Not a prompt. A character.

Every response our characters give runs through three stages: Perception → Cognition (reactive, limbic, or strategic) → Conversational Layers. Here's what that looked like across three real sessions — same architecture, three different outcomes.

PERCEPTION

COGNITION

Reactive

Limbic

Strategic

CONVERSATIONAL LAYERS

Trace 1 — Strategic dominant (session b955728d)

Visitor opened with honesty. Marcus weighed it for 1.2 seconds. No rule fired — he decided.

"Honestly, that's a hell of a way to start a conversation, hermano… Either way, I'm listening…"

Trace 2 — Limbic dominant, let_in=TRUE (session 172908a5)

Visitor made Marcus genuinely laugh. The limbic tier tipped the call. The rope opened.

"Hell of a closer, hermano… That's what gets you in, not the money, not the connections, just you being you and making me laugh. Come on, you're good."

Trace 3 — Reactive alone (session 3c43b985)

Name-drop. Zero deliberation. Reactive tier fired alone this turn. Nine words.

"Who the hell are you to Sarah Chen, kid?"

GET IN TOUCH

Get in touch.

Borys reads every message personally. No CRM, no sequence, no automation.

Rather just meet one of them? — demo.simagentix.xyz → (opens in new tab)