The reasoner orchestrates. The frontier consults.
Eve-Education™ Fusion v5 is the compositional architecture behind ArthurAI™. Three layers cooperate per request: a classifier (Microsoft Phi-3) reads the question, an orchestrator named Eve-Education™ F5/reasoner — a Microsoft Phi-4-derived Small Reasoning Model LoRA fine-tuned on Eve-Genesis™ (Education Edition) — scaffolds the reasoning and delegates sub-problems to three frontier models as consultants, and then synthesises the answer in the discipline's idiom. Engineered at HYVE Labs, the AI/ML R&D division of MindHYVE.ai, Inc..
One classifier. One reasoner. Three frontier consultants.
- CLASSIFIER
Microsoft Phi-3
Routes each request to the right depth — quick lookup versus deep investigation — in under a second. The first responder of the architecture; decides which downstream models compose for the query at hand.
- REASONER
Eve-Education™ F5/reasoner — a Microsoft Phi-4-derived Small Reasoning Model, LoRA fine-tuned on Eve-Genesis™ (Education Edition)
The education reasoner. Trained on a synthetic reasoning dataset built from canonical education methodology — specifically the reasoning modes the discipline actually uses (analogical, Socratic, phenomenological). The orchestrator: scaffolds the reasoning structure, delegates sub-problems to the frontier slots as consultants, synthesises the final answer in the discipline’s idiom. No customer data in training.
- FRONTIER SLOTS
Anthropic Claude Opus 4.7 · OpenAI GPT-5.4 · Meta Llama 4 Scout
Three frontier models composed dynamically per request — Opus for deep synthesis and judging, GPT for alternative reasoning paths, Llama 4 Scout for long-context analysis. Provider-agnostic by design; any model can be swapped without rebuilding the agent. The frontier models work for us, not the other way around.
We don’t compete with frontier labs on foundation-model capability. We compose frontier models into our reasoning systems. Eve-Genesis™ training data is 100% synthetic by construction — no customer conversation is ever used to train a model.
What a request actually does.
The diagram below traces a single request through the architecture. The accent column is the reasoner's work — it scaffolds the response, delegates sub-problems, and synthesises the consultants' output. Two requests entering the same fabric one minute apart may be served by different combinations. The classifier decides; the reasoner orchestrates.
- Question enterssystem
A request enters the fabric. The classifier reads it.
- Classifier routesclassifier
Phi-3 decides the depth and modality. Sub-200ms.
- Reasoner scaffoldsreasoner (ours)
The F5/reasoner builds the reasoning structure — decides what sub-questions to ask, in what order, of which consultants.
- Frontier consultsfrontier consultant
Opus, GPT, Scout receive their assigned sub-problems and return their work. Provider-agnostic; per-request composition.
- Reasoner synthesisesreasoner (ours)
The reasoner receives the consultants’ output and synthesises the final answer in the discipline’s idiom — not the internet’s.
- Answer returnssystem
The response is rendered in the surface that asked for it.
The accented column traces the reasoner's work. The reasoner scaffolds the response, delegates sub-problems to the frontier consultants, and synthesises their output in the discipline's idiom. The frontier models are commodity capabilities the reasoner employs — they are not the entity that owns the request.
Image-bearing requests get a parallel consultant.
Meta Llama 4 Vision — vision route. A parallel routing path for image-bearing requests. The reasoner delegates vision-bearing sub-problems to Llama 4 Vision and receives structured output it can fold back into its synthesis. Open-weight family alignment with Llama 4 Scout.
On Arthur, the vision route handles diagram analysis, handwritten-work review, and document-bearing assessment input. On Chiron, the same architectural slot powers radiology image annotation. Same architecture, vertical-calibrated consultant pool.
What the architecture buys us.
Frontier substitution is cheap
When a new frontier model launches, we evaluate whether it belongs in the rotation. The reasoner is unchanged; the SRM keeps doing its job; the platform is upgraded without disruption. The architecture is structurally hedged against frontier commoditisation.
Regulatory constraints honour-able without rebuild
When a jurisdiction or institution prohibits a particular vendor, we route that institution's traffic around that vendor. The composition is per-request; substitution is operational. No rebuild required.
Cost shape is controllable
The bulk of educational reasoning runs through the Phi-4 SRM at near-Phi-4 cost. Frontier inference is reserved for cases the reasoner judges to warrant it. Unit economics survive Pakistan and East Africa pricing realities.
The IP is ours
The frontier models work for us, not the other way around. The reasoner owns the request. The frontier models are commodity consultants. Anthropic and OpenAI have the consultants; we have the orchestrator that knows what to ask them.
HYVE Labs.
HYVE Labs is the AI, ML, and innovation R&D division of MindHYVE.ai, Inc. The Eve-Genesis methodology, the Eve-Fusion v5 architecture, and the F5/reasoner orchestration pattern were all invented at HYVE Labs.