R&D

Applied research with an obligation to deliver

We work on the border between what technique allows and what regulation demands. Each line is born from a real client problem or a real market failure, and dies if it does not pass its own experiment.

How we work

Five rules of method

The thesis is written before the code

Each line starts with a falsifiable specification: what we claim, the metric that measures it, and the result that would refute it. If it cannot be refuted, it is not research.

Mandatory classical fallback

No experimental capability — quantum, generative or heuristic — enters a product without a deterministic reference path that always runs.

Nothing is emitted unverified

The cortex that proposes is separate from the core that certifies. A result without a verifier is a hypothesis, and is communicated as such.

Recorded before it is run

The hypothesis, the metric and the criterion that would refute it are written down before the experiment is launched, in a log that only accepts additions and is never rewritten. A result that contradicts our own thesis enters the file with the same weight as a favourable one, and several have.

Signed and repeatable

Every delivery is cryptographically signed and every experiment fixes its seed, so that anyone granted access in a due-diligence process can repeat the run and obtain exactly the same result. Runs on third-party hardware are approved one by one, with the quota capped in advance.

Active lines

Seven open fronts

01 · Governed cryptographic agility

Changing algorithm is not enough: you must be able to change it again, hot, without downgrading what is already signed, leaving evidence verifiable years later. This is the shared thesis of Moeris, Éter and Nova.

02 · Sovereign, traceable AI

Models trained in-house on versioned, signed data, runnable locally with no data egress. It answers the regulatory question almost nobody can: exactly what was this model trained on.

03 · Certified computation and structural honesty

Separating a cortex that speculates from a core that only certifies the provable. Active research lives in EVA: competing heterogeneous planes, a monotone blackboard and independent verification before any result is emitted.

04 · Proprietary power-flow solver

A power-flow engine with no commercial dependencies, ten to fifty times faster than the conventional tool, turning grid operation from reactive into anticipatory.

05 · Regulatory knowledge agents

Experts specialised by autonomous region and jurisdiction, able to support a due diligence report or an expert opinion with the specific regulation and case law that applies.

06 · Similarity between genetic conditions

Given the genetic and observed-trait profile of a case, propose which other conditions resemble it and why, with the source cited at every step and without the data leaving the machine it runs on. It does not diagnose and does not replace a clinician: it orders hypotheses so a specialist can rule them out sooner. It is at pilot stage, with a single condition covered end to end.

07 · Applied quantum computing, without the hype

Quantum-ready operators with a mandatory classical fallback from day one: the reference classical algorithm always runs, and the availability of a quantum backend is never presented as demonstrated advantage.

Quantum programme

Four sibling lines, independent physics and IP

They share experimental infrastructure and discipline — hypothesis pre-registered before running, averaging over many seeds, negative results published all the same — but each has its own physics and its own intellectual property line. For each we publish the question, the method and the status; never the mechanism.

Cerberus

Active protection of quantum information

The question: Can a quantum system be defended better by anticipating noise instead of resisting it uniformly all the time?
How it is attacked: A programme of pre-registered numerical experiments over noise models with memory, verified across several independent simulation engines and anchored on real quantum hardware: the circuits sent to the processor are public-domain characterisation sequences, with the quota capped and approved before every submission.
Status: Numerical proof of concept complete and cross-verified, with the reference behaviour measured on a real quantum processor. The next leap — taking our own mechanism to the device — is conditional on protecting the intellectual property first.
What we do not claim: The advantage is demonstrated in rigorous simulation, not on a device. And the hardware has already returned an uncomfortable result that qualifies one of our starting premises: it is recorded in the file with the same weight as the favourable ones, and it conditions the next phase. We also discarded four hypotheses along the way, among them that an AI predictor would improve the result: it does not, and knowing that avoided investing where there was nothing.

Heracles

Detecting what cannot be measured

The question: If part of a system’s energy leaks into a sector we have no access to, could we detect its shadow from the sector we do observe?
How it is attacked: A closed, unitary system where the loss is purely epistemic — it arises from being unable to measure a part, not from an invented sink — with a pre-registered null hypothesis so a false positive is caught before it is celebrated.
Status: Detector validated in simulation: it distinguishes leakage into a non-observable sector from ordinary decoherence and recovers the known parameter, with no false positives under the null hypothesis.
What we do not claim: We do not claim to have proven any exotic physics. We reproduce the situation of an observer with restricted access, which is a legitimate, measurable and falsifiable problem. Anything else would be marketing.

Midas

Defending cryptographic assets against the quantum threat

The question: Can the presence of an active quantum attacker be detected before it executes the mass attack on digital assets?
How it is attacked: Three separate layers: inventory and risk scoring of real public exposure, an honest quantification of the distance to the threat through resource extrapolation, and an early-warning defensive mechanism.
Status: Feasibility, quantified threat model and pre-registered experimental design closed. Detection and defensive mechanism under phased development with hard gates.
What we do not claim: We do not break real cryptography: today that is physically impossible and saying otherwise would be a lie. We do not attack third-party networks or keys. And we do not replace post-quantum cryptography: we complement it with detection and orchestration.

Chronos

Energy cost of computation

The question: How far can the energy a computation dissipates as heat be reduced, operating with reversible logic and slow, controlled evolution?
How it is attacked: A theoretical framework resting on the recognised physical limits of computation, hypotheses graded by level — established physics, theory, conjecture — and a governance rule: no claim rests on conjecture.
Status: Foundations and design closed, with falsifiable hypotheses formulated and explicit success and refutation criteria. Comparative simulation in the next phase.
What we do not claim: It is not a hardware project and it does not claim to create energy: every balance sits inside the first law of thermodynamics. Nor does it promise zero cost: there is a real trade-off between speed and dissipation, and the advantage may exist only in very cold regimes. If so, we will say so.

None of these lines is communicated with numerical results, parameters or mechanisms. The reason is explicit: disclosing a mechanism before securing its priority date destroys any possibility of protecting it. That is why only public-domain characterisation sequences are sent to a third-party processor, while our own mechanism runs locally. What is public is the discipline: if an experiment fails, it is recorded and published.

What we do not claim yet

There is execution on real quantum hardware: several jobs submitted to an IBM processor, using standard characterisation sequences under a pre-registered quota budget. What we do not claim is demonstrated advantage on device: our own mechanism runs locally as a deliberate intellectual-property decision, and having a quantum backend is never, on its own, an advantage. Some of those runs returned a negative result that qualifies one of our starting premises: it is recorded exactly like the favourable results. EVA is at specification stage and its central thesis may not hold; that is why the controlled experiment exists before investing in infrastructure. When something fails, we say so.