• The founding principle
• The four layers
• The processing chain
• Who the foundation serves
• On the topology
• Disclosure policy

The founding principle

Everything rests on one choice: no floating-point operation anywhere in the normative computation path. The arithmetic substrate is integer, fixed-point, with deterministic saturation and rounding. The same input produces exactly the same output, whatever the language, the compiler or the processor: this is what the specification requires, and what three independent implementations, in C, Python and Rust, check by producing the same conformance hash on x86-64 and ARM processors.

In floating point, two machines computing the same thing obtain slightly different results. On a neuron sitting at the edge of its threshold, that difference is enough to make one fire and not the other; one time step later, the two executions have diverged for good. In integer arithmetic there is no difference at all. That is what makes it possible to replay a decision, to compare two trajectories, and to establish that an embedded system behaved as the validated version did.

Determinism serves here as a working instrument, not merely as a final property: mapping an internal state at a given instant, replaying a trajectory from a checkpoint, opening conditional branches, comparing two variants causally from the same starting point, localising an error, resuming after an interruption.

The four layers

The programme is structured like an operating system: a contract between the hardware and what runs on it. Each layer rests on the one below, and each remains usable without those above.

The five deposited documents concern layers 0 to 2; layer 3 is the research frontier, at the prototype stage. An industrial user who needs to replay a decision identically can stop at layer 1 and needs nothing else.

This page describes the technical architecture, which is public. The research method by which I put my hypotheses to the test is a different matter: in my method, AI tools serve as my assistants; it keeps only what a certificate proves, and it remains confidential.

The processing chain

The entire system exists in memory as a directly addressable space; it is designed so that only the relevant part activates at each time step. That is the condition for embeddable execution; the gain in computing time and energy is yet to be measured. The system is designed so that an internal change is first tested separately, its consequences compared against the reference trajectory, then accepted or rejected, without global retraining. After an interruption, a full checkpoint allows exact resumption (checked on a test case: the same final state, bit for bit).

Who the foundation serves

Certification of embedded software

Aerospace, automotive, medical, rail: reproducibility is required by the applicable standards in these sectors, obtained today by freezing hardware and compiler, sometimes for twenty years. The contract provides it by its very definition, including after a port — checked between optimisation levels -O0 and -O2, on two architectures (x86-64 and aarch64). It is not a certification on its own: it supplies one piece of one, the evidence that the computation stays identical.

Porting and hardware migration

Proving that a change of processor does not alter the behaviour of a deployed system is a recurring cost and poorly tooled. The contract reduces it to a comparison of fingerprints.

Scientific reproducibility

The difficulty of reproducing a numerical result affects all of simulation, well beyond neural networks. A substrate on which two laboratories obtain exactly the same result is a cross-cutting tool, independent of the application domain.

On large-scale topology

Some experiments use a candidate topology derived from public connectomics data, as a research test bench. Only the connection structure comes from it: the weights, the dynamics and the learning rules belong to the project, and nothing here is a biologically faithful simulation. The dedicated page describes this testbed.

Disclosure policy

The programme follows a four-level disclosure matrix, applied systematically:

Published results come with their seals, fingerprints and protocols; reserved elements belong to levels 3 and 4. This boundary is declared, stable, and the same for everyone.