Europe should understand its own sensitive text
Aeteos builds Percipion so investigators, analysts and operators can read meaning in language without sending that language, or that judgement, to a third-party model.
The machine should help a human see. It should not speak over them, and it should not create a dependence they cannot unwind.
The problem is no longer volume. It is control
Ten years ago the bottleneck was time. Analysts sat with reports, messages and case files and tried to extract precise meaning before they could act. Ordinary software could store the text. It could not modernise the understanding.
Today the bottleneck is different. In defence, law enforcement and critical organisations, AI and language tools already read at scale. They flag keywords, cluster patterns and raise alerts faster than any team.
They also fail the test that matters here.
Most of those tools sit on foreign technical stacks. They are opaque. They hallucinate. They can be poisoned or steered. They are difficult to inspect when a court, a commander or a regulator asks why. For classified or sensitive text, that is not an IT inconvenience. It is a loss of autonomy.
Therefore, Europe needs a way to understand textual threats that runs on its own infrastructure, can be explained, and leaves the decision with a human.
That is the vision Percipion was built to serve.
Do not imitate intelligence. Reproduce the mechanism
Percipion does not guess the next word. It is symbolic cognitive computing: a machine that represents concepts, context, valence and intent, then reasons over that representation.
That choice comes from how people actually understand language. The work draws on cognitive psychology, neuroscience, linguistics, philosophy and computer science, not as decoration, but as the design method.
A human mind does not only match strings. It builds a model of the situation, including emotion, need and drift, and then decides. If a machine is to help with that work, it needs a comparable structure: a contextual lexicon, knowledge schemas, and an inference that can be inspected.
We do not claim the machine is a mind. We claim it can carry a careful version of the operations a mind uses on text, represent, grasp, infer, so the protector in front of it stays responsible.
Short line, keep it :
We don’t imitate intelligence. We reproduce its fundamental mechanisms, in order to help, serve and protect people.
Delegate the reading. Do not delegate the authority
People have always built tools to carry part of a task. That is legitimate. It is not legitimate when the tool cannot be limited, inspected or switched off.
The operator must still decide when to delegate, over which corpus, and for what purpose. They must understand enough of the mechanism to remain answerable for the outcome. A system that cannot explain itself, or that can only run in someone else’s cloud, slowly takes that answerability away.
Percipion is designed as a subordinate instrument: it reads, represents and proposes. A human accepts, rejects or acts.
What sovereignty means here
Control
You deploy it on your premises or fully offline. You decide what it ingests. Aeteos does not capture your data.
Security
The engine is built without third-party components and without training on your case files. Incoming text cannot fine-tune it, so it cannot be poisoned the way a learning model can.
Compliance
Inferences can be inspected. That is the posture required by the EU AI Act and by NIS 2-era operators who must stay in control of high-risk processing.
Accountability
The assessment is there to support a human decision, or to trigger a response a human has defined. It does not replace the analyst.
The vision, in one line
A sovereign, offline-capable, explainable system that catches weak signals in text by reproducing human-like comprehension, with no third-party dependency, and with the human still in charge.
At Aeteos, that is what we are building.
Evaluate it on your infrastructure
If you want a closed-door demonstration against your own text, write to us.
We do not put a public form on this page. One mailbox, handled by the team.