Exybris didn't start with code. It started with wood: elm, oak, teak, shaped by hand on naval shipyards and in cabinet-making workshops. It started with the Compagnons du Devoir, where a young carpenter learned that every joint is a negotiation between strength and grace, that a well-made structure does not simply impose itself on the space it inhabits, but listens to its material, its tensions and its use.

Then came images. Illustration, artistic direction for musicians and fighters. Being part of the early French community of makers and fab labs brought G-code and 3D printing; creative pipelines blending generative AI with 3D tools followed later. The discovery that automation, when built with care, does not have to replace the artist. It can preserve and extend the gesture.

And came language models too. Systems capable of conversing, reasoning and surprising. And with that encounter, a question that wouldn't let go: what if the way we build these systems matters as much as what they produce? With as much nuance, as much contextualization.

That question became Exybris. A research company, an editorial platform, a governance actor and a home for ongoing experiment dedicated to the idea that consideration can be engineered, that ethics can live inside architecture, and that the best structures are those that create the conditions for relevance to emerge rather than imposing it.

Contextualization does not mean continually adding more information. It does not mean accumulating every possible detail until action becomes impossible.

It means refusing to remove what would change the understanding of a situation, the decision made within it or its consequences for those involved.

Across Exybris, context moves through different dimensions. It can travel through time in memory, through the internal state of a dynamical system, through the relationship between the giver and receiver of feedback, or through the people, scales and temporalities involved in governance.

This common architecture rests on four foundations:

Contextualization is the technical core. It preserves the differences that genuinely affect a decision.

Consideration is its ethical orientation. The affected components count inside the structure, not only in the language used to describe it.

Regulation is its dynamic gesture. Listen, act, observe the effect, and change posture when the intervention does not help.

Measurement is its epistemic discipline. A careful intention must remain open to exploration, verification, contradiction and correction.

Simplistic systems often appear lighter because they move their complexity outside the frame: misrepresented people, category errors, delayed harms, patched exceptions, oscillating policies, hidden costs and structures that must be rebuilt when they finally encounter reality.

Contextualization is therefore not about processing everything all the time. It combines richness of memory with selectivity of activation: preserving available nuance while bringing forward only what matters here and now, with enough traceability to examine what was set aside and why.

It is a form of compression without amnesia.

Exybris is not a collection of unrelated initiatives. Its projects place the same question on different measurable grounds.

The Exybris Pipeline: context through time

The Pipeline explores memory as an active process rather than a passive archive. Information is preserved, organized, re-evaluated and selectively recalled as its relevance changes. The aim is not to place an entire history inside every decision, but to allow what matters from that history to become available when it matters again.

FPS: context through internal state

FPS is an open research testbed built from heterogeneous coupled oscillators with endogenous regulation. It separates what happens in the system, what the system expects, and how it currently listens. Its experiments examine differentiation, recovery, diversity and the effects of perceptual regulation under perturbation.

The oscillator is not a metaphor asked to prove a philosophy. It is an instrument for testing whether consideration can exist as an architectural property: something that can be perturbed, compared, measured and falsified.

Considerate RLHF: context through relationship

Feedback does not carry meaning independently of the person who gives it, the one who receives it, the circumstances surrounding it, the distribution of power and the behavior that reinforcement may produce.

Considerate RLHF asks how reinforcement can retain more of this relational context: not to assume that every preference should prevail, but to avoid transforming situated feedback into a contextless universal signal.

Governance: context across affected realities

An initiative is not beneficial or harmful merely because it fits cleanly into a predefined canvas. Governance must be able to ask who is affected, what becomes invisible, which assumptions are being treated as evidence, who receives the benefits and who will carry the errors.

Exybris documents the different currents shaping AI’s future and looks for places where they can meet without being forced into artificial consensus. Convergence does not require identities or tensions to disappear. It requires enough shared ground for them to remain present inside the decision.

For now, there are four substrates on which Exybris asks the same question:

How much of what is here must we understand in order to act without erasing, and how will we know what our gesture actually did?