I am developing a deterministic, learning and evolvable computing substrate for building artificial intelligence systems able to work over the long term, sometimes without an Internet connection, while keeping a verifiable state and a replayable trace of their decisions.

As artificial intelligence moves into robotics, industry, transport, research and space exploration, understanding, traceability and auditability become essential. When an AI acts on a robot, a machine, a vehicle or a scientific instrument, it must be possible to examine what it did, understand what led to its decision, replay the execution and verify the effect of a change, rather than depend on an entirely opaque way of working.

Starting from the smallest building blocks

I started from the most elementary elements: representation of states, deterministic computation, neurons, synapses, propagation, plasticity, memory and the consequences of actions. I seek to assemble these components step by step without losing the ability to examine each stage.

The project is neuromorphic through one of its development paths, but it does not depend on neuromorphic hardware. Its deterministic substrate can provide building blocks for software, embedded systems, robots, simulations or architectures that interact with language models.

The aim is to allow a system to perceive its environment, build a usable representation of it, learn locally, keep its experience, evolve certain connections or blocks and add new capabilities without necessarily restarting all of its training.

An intelligence designed for embedded use

Frugality is a design property of the project. An embedded AI should not keep its whole network active at all times: it should be able to activate the neurons, synapses and blocks needed for the present situation, leaving the rest of the system at rest. This selective activation aims to limit the computation, memory and hardware resources needed to run it.

The first protocols already establish a sparse and bounded activation: only some neurons and some connections are engaged, propagation is controlled and its computing cost is tracked at every step. These results remain tied to the protocols tested; they are not yet a general measurement of energy savings.

The energy gain remains to be measured on real platforms. What I am already trying to establish is an architecture able to work with constrained resources, to keep only the useful state and to evolve its capabilities without systematically retraining the whole system.

What is already established

The work done has made it possible to develop and test:

• a deterministic, bit-exact arithmetic;
• simulated neurons and synapses with signal propagation;
• a sparse and bounded activation, with controlled propagation and a tracked computing cost;
• plasticity and learning mechanisms;
• an episodic memory able to record, index and replay episodes;
• state preservation and restoration;
• deterministic replay of verified executions;
• analysis of the consequences of an action in a simulated world;
• diagnostic and causal-analysis tools;
• experiments on networks derived from connectome data;
• components tested in scenarios involving language models.

On precise protocols, this work has already established that computations can be reproduced bit for bit, that states can be resumed after an interruption, that circuits can learn simple tasks and that some executions can be audited after the fact. The checks available include 672 identical codes and 48 identical per-step fingerprints in a cross-verification protocol.

These results do not yet amount to general intelligence. They form the experimental foundations needed to measure how far a learning, persistent and verifiable architecture can go.

A parallel line of research with language models

Language models are a complementary line of research. They can propose an intention, a hypothesis, an interpretation or an action. The deterministic substrate can then provide an execution framework, a persistent memory, the observation of consequences and a replayable verification.

The experiments carried out with language models have already produced very good results in several public challenges and research projects. These results can be checked by third parties from leaderboards, official lists, public repositories, certificates and reproducible computations.

They show that an approach combining probabilistic models, structured search, exploration, memory and deterministic verification can produce controllable solutions in varied fields. They are not yet a proof of general intelligence, but they encourage me to pursue research on how probabilistic models and a deterministic substrate fit together.

The Atlas and the challenges serve in particular to study how a system can select what it observes, map the problems it meets and gradually build a usable representation of the world it explores.

My results

Each result is published by the challenge referee, with its source and date; click a card for the details.

Circle packings

Two tables kept by third parties: every packing is exact and can be rechecked.

European Space Agency (ESA)

SpOC 4: four first places on ESA’s post-competition leaderboards. The sum of the four ranks is the best possible (1 + 1 + 1 + 1), separate from the official overall ranking fixed at the deadline.

These submissions were made after 30 June 2026: the leaderboards can change and are distinct from the official competition result. See all four ESA pages →

The fruit fly connectome

Four FlyWire Codex challenges, where my scores are at the top after closing.

Number theory

A puzzle published every month, with its solver list.

Mathematics

Four construction problems whose solutions are checked exactly.

Deposited documents

The foundations, open access on Zenodo.

All the results →

These results were obtained on accessible hardware, detailed in the research programme. Every figure on this site carries its source and its date; its limits are written next to it.

Read the research programme →

Systems that must be able to last

In the long term, an embedded artificial intelligence should be able to work locally, without a permanent connection, keep a useful memory, adapt to new situations, receive new blocks of capability and explain its decisions.

The envisaged applications cover autonomous robotics, industrial machines, embedded systems, vehicles and drones, assistance robots, scientific simulation, autonomous software, safety systems, space exploration and architectures that use language models.

In all these fields the stake is the same: not only to produce an answer, but to be able to establish how the system acted, under what conditions, with what internal state and with what consequences.

An ambition to be measured step by step

My aim is to build a model that is safe, certifiable and auditable. Every step is recorded: hypotheses, versions, protocols, results, failures, corrections and limits.

Artificial general intelligence is the horizon of this research, not a result already claimed. The current work consists in establishing, step by step, whether a system can be learning, persistent, modular, adaptive, usable offline and auditable, without its internal architecture having to be made entirely public.