• Vision and aim
• A deterministic substrate with general scope
• Frugality and selective activation
• What has already been achieved
• The results with language models
• The role of language models and of the Atlas
• Applications
• Certification and the European AI regulation
• References

Vision and aim

I am developing a deterministic, learning and evolvable computing substrate intended for building embedded artificial intelligence systems that are safe, verifiable and auditable.

My aim is to create a model whose properties can be established by proofs, tests and execution traces rather than by a mere promise that it works. I started from the most elementary building blocks: representation of states, deterministic computation, neurons, synapses, propagation, plasticity, memory and the consequences of actions. I seek to assemble these components step by step to form a richer system, without losing the ability to examine each stage.

The development is recorded over time: hypotheses, versions, protocols, positive results, failures, corrections, limits and governance decisions. This documentary continuity makes it possible to know what has actually been tested, under what conditions, with what result and with what degree of confidence.

As artificial intelligence enters everyday life, robotics, industry, research and space exploration, understanding, traceability and auditability become essential. When a system acts on a robot, a machine, a vehicle or a scientific instrument, it must be possible to examine what it did, verify what led to its action and reproduce its behaviour after the fact, rather than depend on an entirely opaque way of working.

An embedded AI must be able to work under constrained conditions, sometimes without a network connection, while keeping a coherent state and a memory of its experiences. In an industrial plant, an autonomous robot or a space mission, an error cannot be corrected only by a new probabilistic answer: it must be possible to locate the cause, replay the execution and demonstrate that a change really improves the system.

A deterministic substrate with general scope

The project is neuromorphic through one of its development paths. It makes it possible to study simulated neurons and synapses, signal propagation, the activation and inhibition of units, plasticity and memory.

But it is not limited to neuromorphic hardware. Its starting point is a deterministic substrate able to evolve towards a neuromorphic architecture. Thanks to determinism, its building blocks can be used in different software, hardware and environments: robotics, industry, autonomous systems, embedded software, scientific simulation or interaction with language models.

The same execution must be able to produce the same result under the intended conditions, bit for bit. States can be recorded, restored and replayed. This property makes it possible to examine an execution after the fact, to compare two versions of a system, to locate a divergence and to verify the real effect of a change.

Frugality and selective activation

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.

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. This selective activation aims to limit the computation, memory and hardware resources needed to run it.

These results remain tied to the protocols tested. They are not yet a general measurement of energy savings. This gain will have to be measured on real platforms.

What has already been achieved

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

• a deterministic arithmetic substrate;
• circuits of simulated neurons and synapses;
• signal propagation in structured networks;
• a sparse and bounded activation, with controlled propagation and a tracked computing cost;
• plasticity mechanisms;
• an episodic memory able to record, index and replay episodes;
• state preservation and restoration mechanisms;
• bit-exact replay on the verified protocols;
• continuity checks and resumption after interruption;
• an analysis of the consequences of an action in a simulated world;
• causal-analysis and diagnostic tools;
• an internal state representation;
• environments for testing learning, memory and adaptation;
• experiments on networks derived from connectome data;
• tests of some components in scenarios involving language models.

This work has already made it possible to establish several properties on precise protocols: computations can be reproduced bit for bit, states can be resumed after an interruption, circuits can learn simple tasks and some executions can be audited after the fact.

Cross-checks have also been carried out on the implementations and the execution fingerprints. The results available show an exact match on the protocols tested, with 672 identical codes and 48 identical per-step fingerprints in one of the cross-verification checks.

These results do not mean that general intelligence has been reached. They establish technical and experimental foundations: verifiable computation, persistent state, local learning, plasticity, memory, controlled activation and replayable execution.

The essential next step is to demonstrate that memory and the consequences of an experience really change a later decision in the complete nominal path. This property will have to be measured with controlled experiments and ablations, in order to tell a mere recorded trace from a memory that truly acts on behaviour.

The results with language models

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. The results do not rest only on an internal claim by the system: they are confronted with verifiers, public platforms or independent criteria.

These experiments are not yet a proof of general intelligence and do not mean that a language model fully understands its own functioning. They do show, however, that the building blocks developed can be used effectively in the environment of a probabilistic model and produce results that withstand outside verification.

This public validation encourages me to pursue research on how probabilistic models and a deterministic substrate fit together.

The role of language models and of the Atlas

Language models are a parallel line of research. They can propose an intention, a hypothesis, an interpretation or an action, but they do not necessarily provide on their own a persistent causal memory, a local plasticity, an explicitly verifiable representation or an exact replay of their internal execution.

I therefore test certain building blocks of the project in scenarios involving probabilistic models. The question is not to ask the language model to explain its own functioning perfectly. The question is whether a distinct substrate can receive a probabilistic proposal, execute it in a controlled framework, observe its consequence, keep that experience and compare the following decisions.

The challenges and the Atlas are part of this research. They make it possible to explore how a system can select what it observes, map the problems it meets and gradually build a usable representation of the world it explores.

Applications

Such an architecture could provide a basis for several kinds of systems:

• autonomous robotics;
• industrial machines;
• offline embedded systems;
• vehicles and drones;
• assistance robots;
• scientific simulation;
• autonomous software;
• safety systems;
• controlled interaction with language models;
• critical industrial systems;
• evolvable architectures;
• space exploration.

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.

The project thus aims at an artificial intelligence that is not only able to produce an answer, but able to work over the long term, learn from its environment, keep continuity and make its behaviour examinable.

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, frugal, usable offline and auditable, without its internal architecture having to be made entirely public.

Certification and the European AI regulation

Certifying has a precise meaning here: establishing by computation, with a certificate anyone can recheck, that a result is exact, or that a network reduced in precision and installed on a chip makes the same decisions as the validated model. Every decision replays identically, so it can be examined after the fact.

This matches what regulators ask for. The European AI regulation (Regulation (EU) 2024/1689, the “AI Act”, official text) provides that high-risk AI systems must technically allow the automatic recording of events (logs) over the lifetime of the system (Article 12), must have technical documentation (Article 11) and must be open to human oversight (Article 14). Under the calendar amended in 2026, these obligations apply from 2 December 2027 to stand-alone high-risk systems and from 2 August 2028 to those embedded in regulated products. In aviation, EASA already asks for a check that reducing a network’s precision does not change its decisions. My AI is designed to produce such traces, replayable bit for bit. I claim no conformity with these texts: it is a direction of work, and an official certification is a matter for accredited bodies.

References

1. J.-R. Denoual, A Normative Arithmetic Substrate for Bit-Exact Spiking Neural Networks (ArithSpec), Zenodo, 2026. DOI 10.5281/zenodo.20723008
2. J.-R. Denoual, Per-Tick Certification of Spike-Exact Fixed-Point Spiking Neural Networks, Zenodo, 2026. DOI 10.5281/zenodo.21749446
3. J.-R. Denoual, Internal Margins Collapse, Decision Margins Grow: Why Trained Spiking Networks Tolerate Low Precision, Zenodo, 2026. DOI 10.5281/zenodo.21750036
4. J.-R. Denoual, Rate-Level Robustness in Deterministic Spiking Networks: When Spike-Train Fidelity Overstates Observed Functional Fragility, Zenodo, 2026. DOI 10.5281/zenodo.21750042
5. J.-R. Denoual, Rate-Level Robustness under Quantization in Connectome-Instantiated LIF Models, Zenodo, 2026. DOI 10.5281/zenodo.21750053
6. S. Dorkenwald et al., “Neuronal wiring diagram of an adult brain”, Nature 634, 2024; P. Schlegel et al., “Whole-brain annotation and multi-connectome cell typing of Drosophila”, Nature 634, 2024; FlyWire data, version 783, used for research purposes under their terms of use. codex.flywire.ai (accessed 7 Oct 2026)
7. E. Specht, Packomania “csqv” table. packomania.com (accessed 4 Oct 2026)
8. European Space Agency, SpOC 4 challenge, ESA Optimize platform. github.com/esa/SpOC4 (accessed 4 Oct 2026)
9. FlyWire Codex, Max Quasi-Cliques challenge. codex.flywire.ai (accessed 7 Oct 2026)
10. IBM Research, Ponder This, September 2026. research.ibm.com (accessed 11 Oct 2026)
11. EASA, Artificial Intelligence Concept Paper, Issue 02, March 2024. easa.europa.eu (accessed 3 Oct 2026)
12. E. Friedman, Packing Center, circles in a circle and in a rectangle. erich-friedman.github.io (accessed 11 Oct 2026)
13. European Space Agency, SpOC 4 challenge, Luna Tomato Logistics, ESA Optimize platform. optimize.esa.int (accessed 11 Oct 2026)
14. FlyWire Codex, Minimum Feedback Arc Set and Visual Columns challenges. codex.flywire.ai/app/mfas_challenge; codex.flywire.ai/app/visual_columns_challenge (accessed 7 Oct 2026)
15. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence, Articles 11, 12 and 14. eur-lex.europa.eu (accessed 14 Oct 2026)