System design
Architecture you can inspect
The same layered model shown on the landing page, expanded for technical readers. Each layer maps to a coherent set of capabilities behind one proof culture.
Architecture
Layered. Coherent. Inspectable.
Midcore stacks deterministic execution below adaptive intelligence — so "move fast" never means "skip proof." Expand each layer to see what you get.
Atomic I/O, model routing, offline enforcement, and runtime adapters.
- ▹Atomic writes & crash-safe checkpoints
- ▹Multi-provider LLM routing & harness
- ▹Offline-first policy enforcement
Conceptual request path
Not every deployment enables every stage; this is the maximal path through the autonomy and delivery surfaces.
- Ingress — prompts, tools, and policy context enter the prompt harness.
- Planning & tools — agentic loop, policy-aware tool routing, and sandboxed execution.
- Intelligence services — retrieval, memory, compaction, and ranking.
- Operational layer — autonomy levels, streaming, hooks, and human gates.
- Outcome Compiler — contracts, gates, evidence ledger, capability passports, packaging.
- Robotics / PubLE — fleet registry, brain pipeline (sense / think / act / learn), safety envelopes, telemetry, simulation, and training exports with the same evidence shape.
- Scientific research — Scout / Simulation / Survey / Surveillance tracks, wizard-driven studies, multi-scenario bundles, and resumable long runs under shared identity and billing.