Composition, not stack.

"Stack" is the wrong word for what we have built. A stack is a fixed set of layers. A request enters at the top, passes through every layer, exits at the bottom. The architecture is determined; the layers are persistent; the change cost is high. That is not what Eve-Education™ Fusion v5 is.

What we call a compositional fabric is dynamic. A request enters; a classifier reads it; a composition is chosen — which models cooperate on this particular request, in which sequence, with which weights. Two requests entering the same fabric one minute apart may be served by different model combinations. The composition is per-request.

A typical Eve-Education^™ request is composed of five cooperating models plus an optional vision route. They are:

• Microsoft Phi-3 — Classifier. Sub-200ms routing across workload type, modality, reasoning intensity, and context-length needs.
• Microsoft Phi-4 — Education reasoner. LoRA-fine-tuned on Eve-Genesis (Education Edition). The domain-specialized workhorse for the bulk of educational reasoning.
• Anthropic Claude Opus 4.7 — Frontier reasoning. Best-in-class long-form structured reasoning. Used when the case is high-stakes and nuance matters.
• OpenAI GPT-5.4 — Frontier reasoning. Frontier general reasoning. Complementary to Opus for cross-validation and ensemble routing.
• Meta Llama 4 Scout — Longitudinal context. Long-context reasoning across a learner’s complete educational record in a single thought.

That is one configuration. It is not a fixed schema. Some requests pass through only the classifier and the SRM and exit. Some requests are routed to all three frontier slots simultaneously and the outputs are reconciled. Some requests are routed to the vision model in addition to a reasoning path. The classifier decides.

Three things happen when an architecture is compositional rather than stacked.

First, frontier substitution is cheap. When a new frontier-grade model is released, the question for us is not "rewrite the platform on the new model" but "evaluate whether the new model belongs in the frontier rotation for which workloads." The classifier learns to route to it; the SRM keeps doing its job; the platform is upgraded without disruption. We have replaced frontier slots three times since the architecture stabilised. Nothing about the product had to change.

Second, regulatory constraints can be honoured without rebuilding. When a jurisdiction or institution prohibits a particular vendor, we can route that institution's traffic around that vendor. The compositional fabric is what makes this possible. A stack baked to a particular vendor cannot. We have institutions for whom no traffic ever touches a particular frontier provider. The product still works.

Third, 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 classifier judges to warrant it. The result is unit economics that survive deployments in markets where frontier-only architectures cannot. That is, again, the access race.

Calling the architecture a stack invites a particular set of design mistakes — fixed layers, fixed contracts, fixed costs. Calling it a compositional fabric invites the correct set of design questions — what is the classifier reading, which models cooperate on this workload, what is the reconciliation pattern when they disagree. The metaphor sets the design conversation.

The same is true outside engineering. When we tell a partner that the architecture is a compositional fabric, they hear "you can route around our compliance constraints," "you can substitute a model that has lost reputational standing," "you can keep the platform working as the frontier moves." That is what they should hear. It is also what is true.

A compositional fabric is not a router with shiny copy. The classifier is doing actual work. The SRM is fine-tuned on Eve-Genesis (Education Edition) — a domain-specific synthetic reasoning corpus that nobody else has. The frontier slots are graded for cooperation patterns, not just raw capability. The composition is engineering, not marketing-language.

The how-it-works gallery is the operating proof. Each of the ten capability deep-dives traces directly back to a place in the platform where the compositional architecture is doing something a single-model architecture could not have done. Lesson generation cooperates across three or more model calls per lesson; the tutor cooperates with the lesson context in real time; the LCP composition runs against age, language, and learning-support adaptation. Stack is the wrong word. Compositional fabric is the right one.