Architecture / Research
The architecture explorer documents boundaries and trade-offs. Every diagram is conceptual unless an entry carries verified implementation evidence.
Text alternative: artifacts enter through a loader and repository boundary; validated metadata informs memory and compute planning; an abstract backend executes work; result decoding and observability leave through explicit application boundaries.
System context
Users, model repositories, local or remote artifact storage, native or browser hosts, compute devices, and observability systems sit outside the core runtime boundary.
Core runtime layers
Loader and initialization, memory and resources, compute and dispatch, and results and observability form four independently testable layers.
Ports and adapters
Traits express the behavior required by the core. Concrete repositories, caches, hardware backends, and telemetry exporters adapt external systems.
Loader and model repository boundary
The loader accepts explicit metadata, source revisions, integrity evidence, and artifact streams. It rejects ambiguity rather than guessing.
Backend abstraction
Backend selection records supported operators, precision, resource limits, and failure modes. Static specialization and dynamic application boundaries are separate decisions.
Memory lifecycle
Host and device allocations, model residency, working buffers, reuse, mapping, eviction, and destruction are represented as explicit lifecycles.
Error topology
Artifact, validation, resource, backend, execution, decoding, and policy failures remain distinguishable through typed errors.
Observability flow
Events, metrics, traces, benchmark identifiers, and source revisions propagate without leaking unapproved user data.
Native execution path
Native execution may use operating-system threads, filesystem-backed artifacts, platform accelerators, and native observability integrations.
Browser-oriented execution path
Browser research must account for workers, sandboxed storage, memory ceilings, WebGPU availability, cross-origin isolation, and browser-specific failure behavior.
Security boundaries
Untrusted artifacts, parsers, unsafe code, build-time dependencies, external sources, and uploaded raw data are treated as distinct trust boundaries.
Extension points
New formats, repositories, backends, schedulers, decoders, and telemetry adapters should enter through focused contracts with documented invariants.