Built from the world,
not from words
Today's AI is fluent, and frozen. It predicts words brilliantly and stops learning the moment training ends. The physical world does not work like text: it is continuous, high-dimensional and noisy, much of it cannot be predicted at all, and it changes underneath the system every day.
Intelligence that models the world before it acts on it
Qualion Intelligence builds world models: systems that learn abstract representations of real-world sensor data, keep what can be predicted and let go of what cannot, and make their predictions in that representation space rather than generating every pixel and token. This is the direction set out in Yann LeCun's programme for autonomous machine intelligence, and the one the field's leading laboratories have now turned to. Our bet is on what the field is not yet doing: systems designed to keep learning after deployment, to stay efficient, and to remain auditable while they do.
Representation, not generation
The encoder learns from sensors, machines and environments what is regular in the world and discards the unpredictable detail. Prediction happens in representation space, which is where the structure of the world lives.
Action-conditioned
The world model predicts how each action will move the environment. Before the system acts, it rolls its options forward, scores them, and chooses a sequence that accomplishes the task inside safety guardrails.
Developmental, not frozen
Capabilities emerge and grow through interaction with the environment, the way development works in nature. The system keeps learning from the world it acts in after deployment, instead of waiting for the next release.
Auditable by construction
Every output traces back to a specific computation that can be inspected, not to a filter bolted on afterwards. When the system is wrong, the error is visible, attributable and fixable rather than fluent and hidden.
An intelligence the physical world can trust
Where a wrong action has physical consequences, probably right is not sufficient. Industrial process control, automation, wearables, robotics and healthcare need systems that are controllable, auditable and safe by design, and regulation increasingly says so: the EU AI Act's high-risk obligations set requirements for traceability, logging and oversight that this architecture is built to meet natively. Trust is the price of entry, and we build for that standard.
An agent with a world model can set its own goals. We verify them.
A system that models the world and plans over that model will, sooner or later, surface objectives nobody authored. Our research programme asks how anyone, including the agent, can know that such an objective is real rather than a convenient story. The answer is a pre-registered, proposer-agnostic protocol that tries to falsify the objective, the same controls run live on a deployed substrate, and optimising adversaries pointed at the verifier so the rate at which it can be fooled is measured rather than assumed.
How do you know a self-created objective is real?
Three manuscripts, released as one package: a pre-registered falsification protocol for self-created objectives, its deployment on a live substrate with blind external forecasts, and a measured adversarial false-accept rate.
Read the researchThree manuscripts, released as one package
Paper I carries the protocol and its adversarial characterisation. Paper II deploys it on a live substrate with blind external forecasts. The letter issues the headline measurement on its own, for a methods audience. Every constant, threshold and predicted verdict is hash-locked before the corresponding run, so nothing can be adjusted after the fact.
The protocol
Verifying Self-Created Objectives I: A Proposer-Agnostic, Pre-Registered Falsification Protocol with a Measured Adversarial False-Accept Rate
PDFThe deployment
Verifying Self-Created Objectives II: Endogenous Self-Extension on a Deployed Substrate, a Pre-Registered Gate, Blind External Forecasts, and a Discovery
PDFThe measurement
A Measured False-Accept Rate for a Proposer-Agnostic Falsification Battery
PDFTwo engineering roles, several openings each
The programme is preparing its system for an enterprise pilot in a regulated sector. Both roles exist to make that transition possible. They are full-time, based in Ljubljana or anywhere in the EU with a regular monthly on-site presence.
ML Research Engineer
Takes the research codebase from internal use to a system that survives external evaluation: evaluation infrastructure, deployment engineering, reproducibility tooling and adversarial testing, working alongside the founder.
AI Safety and Compliance Engineer
Translates the system's auditable properties into documentation that survives external audit under the EU AI Act, ISO/IEC 42001 and sector-specific standards.
The architecture is described in published research and reproducible artefacts, because claims about reliability mean nothing unless others can examine, replicate and challenge them. We build with industry partners, with regulators, and with the research community. If your work depends on systems that must be right, not just plausible, talk to us.