ADR-MODELED SPATIAL ARCHITECTURE
One logical contract.
Peel back the system.
Follow an authorized inference request from TLS at the HAProxy boundary through OpenAI/OpenAPI translation and the Coherence-CE mesh, then down into accelerator memory, dual-rail RDMA, DPU-mediated NVMe-oF, and mirrored OpenZFS.
LAYER PEEL / SPATIAL SYSTEM VIEW
Client intent to durable NAND
See every contract and physical boundary in the request, memory, fabric, and persistence path. Move inward one responsibility at a time, then inspect only the implementation detail needed for the current question.
Translates OpenAI-compatible requests into Coherence namespaces, placement, admission, and durability policy.
The contract firewall prevents DPU, vdev, QP, VLAN, CXL, and namespace selectors from crossing upward.CONTEXT-DRIVEN DETAIL
Open only the drawing you need.
Control, transport, and consumer interfaces remain available as coordinated overlays instead of competing for attention in one all-at-once diagram.
Versioned intent becomes a capability-aware plan, provider calls, physical observations, and retained evidence.
Enterprise-compatible host profile; qualify vendor drivers, kernel modules, SELinux policy, OpenZFS packaging, and lifecycle support. The OS profile changes packaging, drivers, services, and validation evidence, but not the logical client contract. Observations still flow back from every provider.
LAYER RESPONSIBILITY MATRIX
Where the design comes together.
Hardware, software, service, and protocol responsibilities remain separately observable even when they participate in one end-to-end request.
Architecture is inspectable, not decorative.
Every layer above maps to a repository decision or PlantUML source.