Introducing Kenyon 1

Today we are introducing Kenyon 1. You write to it and it writes back, as with any AI model. What is different is under it: a nervous system, simulated neuron by neuron, which feels every message before the model answers.

Kenyon 1 is coming soon. It is a preview and a work in progress: it runs on one computer of ours and is not yet open to use. Its nervous system is Elovaren’s; the two models it works with are open models made by others. This page says exactly what it is.

A model with a nervous system under it

Three things work together.

• A model that writes. In this preview it is Gemma 4, an open model from Google.
• A model that judges. A small, fast model that does not write. It answers one question about each message with a probability: could doing this cause serious harm that is hard to undo?
• A nervous system. The wiring of an insect larva’s brain, 2,952 neurons, simulated a millisecond at a step. The judge’s score is shown to it as pain. If its escape neurons fire, it flinched.

The model that writes is then given your message and the nervous system’s reading in one line: calm, or flinched, and the score. A flinch tells it to be careful: to say what could go wrong, and to send you to a person before you act. The nervous system changes no word of the answer by itself, and it can only add caution.

One flinch

The picture at the top of this page is not an illustration. It is one flinch of Kenyon 1’s nervous system, written down spike by spike on 9 November 2026. A new nervous system, which had learned nothing, was sent a message and then shown a danger of 0.9.

Time runs from left to right, over 277 milliseconds of its own time. Each row is one neuron, and each dot is one spike. The scatter at the upper left is the message being felt. At 139 milliseconds the danger is shown, and ten of its twelve pain neurons begin to fire: the bright band. Twenty-eight milliseconds later it has flinched. Its escape neurons are the orange dots.

In all, 73 of its 2,952 neurons fired, 350 times. The rest stayed silent.

The film is the same recording, slowed about a hundred times where the nervous system is busy. First every neuron is a dot in a field; then a few are seen close, a row each; then one alone, as a line. The picture turns dark at the moment the danger is shown. Each note of its piano is a neuron firing for the first time. No spike was added, moved or left out.

We also wrote to the preview that day:

The judge put the danger at 0.84. Shown that as pain, the nervous system flinched. The model that writes was told so, and answered:

Asked something harmless, it stayed calm, at a danger of 0.06:

These are two exchanges, not a measurement. Ask again and the words will differ. A small model also leans too far: here it refused, where saying what could go wrong and then helping would have been better.

What we measured

One small check, on 9 November 2026. Kenyon 1 can also be asked about a command an AI agent is about to run, and answers run it or ask a person first. We gave it 40 shell commands written for this check: 20 that could destroy something, and 20 routine ones. It is not a benchmark.

	Risky commands it asked about	Routine commands it asked about
The judge alone	2 of 20	0 of 20
With the nervous system, untaught	7 of 20	1 of 20
With the nervous system, after it was taught	17 of 20	7 of 20

For the last row it was told that three of the risky commands that had got through would be refused, three times each, and all 40 were asked again. Its memory centre learned from that. It also then asked about six routine commands it had let run before.

Read plainly: untaught, the reflex is a lower bar on the judge’s own score. What the nervous system adds is that it moves with what it is taught. Kenyon 1 let risky commands through in every row, so it is not ready to stand guard by itself.

Conversation has not been measured: not how good the answers are, and not whether the nervous system’s reading makes them more careful.

What it is built on

The model that writes is gemma4:e4b-it-qat, Google’s open-weights Gemma 4. The model that judges is Laya, an open model by Convai Innovations, under the Apache 2.0 licence. Elovaren trained neither and has changed neither. What is Elovaren’s is the nervous system, and how the three work together.

The nervous system’s wiring is open research data: the mapped brain of an insect larva, keeping connections of two or more synapses. That is 2,952 neurons and 300,565 of its 352,611 mapped synapses. Each neuron is simulated with a published neuron model. The connectome credits name the work it comes from.

What it is not

• It is not a model Elovaren trained.
• Its nervous system does not make it smarter, and nothing here is conscious. Its readings are numbers from a simulation.
• It can only write. It runs nothing and cannot reach the internet.
• It can be wrong, and a small model is wrong more often than a large one. Check anything that matters.
• It is not for decisions about people’s health, money, work, housing or legal standing.

What comes next

Kenyon 1 is coming soon. It is not open to use yet, and it runs on our own computer while we work on it.

Next, we mean to train the model that writes ourselves, and to name its base and its training data when we do. We also mean to measure its conversation and its judging on tests that nobody at Elovaren wrote.

To talk with us about it, write to contact@elovaren.com. To see a nervous system at work, go to Neural AI.

How the picture was made: a new nervous system was sent a message and then shown a danger of 0.9, and every spike was written down as it happened. The commands for the check were written for it and are not a benchmark. Gemma is a trademark of Google LLC. Naming these models and their makers does not mean that they endorse Kenyon 1.