The model reasons.
Your knowledge system knows.
Your policies decide.
Verificate Hybrid Intelligence combines capable open models with cited knowledge, governed decision logic, and controlled inference — so organisations can build AI that is useful, explainable, sovereign, and continuously improvable.
General LLMs do not own your facts, your provenance, or your decision logic.
A large language model is a model of the internet's text. It reasons fluently — but it does not know your entities, cannot cite your records, and should not own your operational decisions. Hybrid Intelligence separates the three jobs so each is done by the layer built for it: general reasoning, cited knowledge, and governed policy.
Five layers. One accountable system.
Open / general model
Language generation and general reasoning — a capable open model you rent or run, never the owner of your facts or your decisions.
Cited EAV knowledge substrate
Facts stored as entity–attribute–value records with source, provenance, licence and time attached — retrievable, updateable, and citable down to the record.
Decision Transformer / specialist policy
Repeatable domain decision logic and constraints, trained on your data and outcomes — deterministic, auditable, and separate from the language layer.
Verificate Inference Engine (HELIX)
Controlled serving on infrastructure you choose, with a confidence score for routing — deliver, confirm, escalate, or block.
Verificate Gate
Quality control at risky AI/agent delivery transitions — the same hard gate that protects AI-written software.
“Hybrid LLM”, “EAV”, “knowledge graph” and “Decision Transformer” are the architecture terms inside this solution. The separation is the point: facts live in a governed substrate, decisions live in a governed policy layer, and the language model narrates — it does not decide. This does not make hallucination impossible; it makes every claim and every decision traceable to a layer you control.
RAG injects context. Hybrid Intelligence governs it.
Retrieval-augmented generation is a legitimate technique: relevant snippets are pasted into the prompt and a general model writes an answer. The precise difference is what is not there: a governed record of which fact drove which answer, a licence and provenance trail, and a decision layer that behaves the same way twice.
The answer changes every time
Ask the same context-injection system the same question twice and you often get a different answer — probability models sample words. For policy, credit, compliance, or student support, that is hard to run a process on.
No clear audit trail
A RAG system may have looked at a document, but the answer rarely records which policy, clause, or record drove the outcome. Reconstructing a chat is not reading a decision record.
Agent bias — measured in study
LLM advisory agents tend to over-act: in our published study a zero-shot frontier agent recommended intervention 43 percentage points more often than the correct institutional policy. A policy trained on the institution's own data did not.
| Context injection — RAG | Hybrid Intelligence | |
|---|---|---|
| Where facts live | In the prompt window, at question time | In a cited, versioned knowledge substrate |
| Provenance & licence | Rarely recorded per answer | Attached to every record, cited per answer |
| Who decides | The general language model | A governed policy layer — the LLM narrates |
| Same input tomorrow | Often a different answer | Same decision — deterministic policy layer |
| Audit trail | Chat log; weak link to policy | Per-decision lineage — what records drove it |
| Where it runs | Often a public API | Your environment — sovereign by design |
Built and exercised at scale — as evidence, not marketing.
Kevin — sovereign knowledge at national scale
EPKevin is Verificate's internal evidence instance of this architecture: an Australian sovereign-knowledge system built as a cited EAV graph over public Australian sources, with licence and provenance carried on every record and relationship edges linking law, government, research and culture. It demonstrates the substrate layer at national scale. Kevin is an architecture and evidence programme — not a live public product today. R
OmniTX — biomedical application of the pattern
EPOmniTX applies the layered architecture in a biomedical/genomics setting: cited domain knowledge and governed decision logic over an open reasoning model. Published as architecture evidence only — availability and results statements follow the claim-substantiation process on the Legal page.
The decision layer is measured, not asserted.
Over-intervention eliminated
A zero-shot frontier agent recommended action 43 percentage points more often than the correct institutional policy; a Decision Transformer trained on the institution's own data did not (EAV-DT study, OULAD). M
Decision flip rate
The trained policy produces the same action for the same complete state — an architectural property, measured at 0% flip rate. M
Learning from outcomes
On a real 251k-case procurement log (BPI-2019), the continually retrained policy beat a frozen one (p≈0.009); a label-permutation placebo collapsed the gain ~389×. M
Sources: EAV-DT paper (arXiv:2606.29280) · CORTEX paper (arXiv:2602.17691) · Research & reproduction
Where cited knowledge and governed decisions matter most.
Regulated knowledge
Answers that must cite the governing record — policy, legislation, contracts — with licence and provenance attached.
Biomedical & technical evidence
Domain corpora where a fluent guess is worse than no answer, and every claim needs a source.
Sovereign & public-sector information
Jurisdiction-first knowledge bases served on infrastructure you control, under licences you can honour.
Decision support
Operational decisions owned by a governed, deterministic policy layer — with the language model narrating, not deciding.
Build AI your organisation can stand behind.
Bring a knowledge domain or a decision process. We'll map it to the layers — model, knowledge substrate, policy, serving, gate — and show you what governed AI looks like on your own infrastructure.