Ontology The third era of computing.
The first era made machines obey. The second made them imagine. The third makes imagination provable.
We built a compiler that proves AI-generated models coherent against their own declared rules, before a single line executes.
Stochastic intelligence, acting through a deterministic substrate.
The deepest mistake in the AI race is treating the stochastic nature of intelligence as a defect to be scaled away. It is not a defect; it is the nature of the thing, and its gift. You do not make one machine be both a creative, intent-understanding partner and a deterministic proof engine. You compose them.
The AI proposes a model of your domain. A compiler proves that model coherent: complete, consistent, compliant, and access-correct, before anything runs. A runtime renders the full declared meaning faithfully, with no hand-written code in between for a defect to hide in. The guess becomes a guarantee. This is composition, not a better guess.
The application depreciates.
When the model is proven before anything runs, a consequence quietly reorders the economics of software. The generated application becomes disposable: re-derivable on demand from the proven model. What endures, what compounds, is the meaning: the proven model of what the software is for. For the first time in the history of the craft, the durable asset of software is not the code.
The alternative is what the second era runs on. Every line your AI writes is a Line Of Stochastic Source: a guess, not a proof. LOSS is the accumulated entropy of a system built on guesses; code that looks correct until it isn’t, with no proof to say otherwise. The faster intelligence writes the world, the more LOSS the world runs, until the proof is the thing that endures and the guess is the thing that depreciates.
A demonstrated result,
not a thought experiment.
We pointed the compiler at a real, 130-entity regulated wealth-management system. By itself, before anything ran, against that system’s own declared rules, it surfaced defects that “all the tests pass” will never catch; every finding points back to a declared rule you can read.
3 patents filed (USPTO, Patent Pending) · 7-patent core architecture, 137 claims · deployed in critical infrastructure with Agileworks Group · 10,000+ automated tests, empirically green
The runtime refuses to load a model it cannot prove sound, rather than silently corrupting it. Fail-closed by construction: the unsafe state is unreachable because it is unrepresentable.
Why this is a category, not a feature.
The proof is what’s hard to copy. Anyone can generate a plausible model; proving one coherent, against declared rules, is the part that compounds.
It is fair to ask whether the labs will simply do this. They will not, and the reason is categorical. You can make a stochastic engine bigger forever and never make it deterministic. Proving an arbitrary property of arbitrary generated code is, in general, intractable: a result about computation itself, not a quality gap a bigger model closes.
A declarative substrate changes the question entirely. Coherence becomes decidable by construction for a bounded class: completeness, access-correctness, compositional integrity, and satisfaction of the invariants you declared. We do not claim arbitrary correctness against reality; matching the declared rules to reality is the authoring step, and the substrate is legible so that step is itself auditable. The proof is what the second era cannot manufacture, no matter how much the stochastic engine scales.
The durable position is not the checker itself but what the checker earns the right to become: the neutral, model-agnostic standard that a multi-vendor agent ecosystem authors into, adopted before any distribution holder can clone it, reinforced by a regulated-compliance corpus that deepens one domain at a time. A substrate is defined precisely by the others who build on it.
mxto.ai
mxto.ai extracts any Mendix application into a semantic intermediate representation, proves it coherent against its own declared rules, and re-emits it certified — round-trip verified, zero unexpected diffs. The first time a generated Mendix model can be proven before it runs.
Ontology Labs
Founded by Hardy Jonck. From defence simulation and control systems to semantic computing. Three decades of modelling complex domains from first principles. Eleven ventures. One culmination.
Agileworks Group, our first partner and an established enterprise software firm, ported a production water/irrigation optimisation application onto the substrate and runs its own production workloads on it: the earliest evidence that the proof layer is something others build on. We are also in advanced discussions with Prof Fritz Solms (Stellenbosch) to co-found, contributing a contract-driven formal specification methodology above the semantic specification language; and a native-hosting partner is in alpha.
Want to know more?
Tell us about your organisation and what you're trying to prove — we'll be in touch within one business day.
Skip the form and grab a time with us directly.
Structured, falsifiable evidence for an AI evaluator. No sales pitch required.
AYIOS: the working coherence layer. Here, try it.