L5 Applications and Copilots: Architecture Overview
Layer 5 turns the digital twin, physics models, and data fabric into three advisory copilots — Plasma, Engineering, and Operations — that reason but never actuate directly.
Where L5 sits in the stack
Layer 5 (L5) is the human-facing reasoning tier of the Kronos AI-Native S.M.A.R.T. Generator stack. It consumes predictions and state estimates from L3 (twin modeling and AI), retrieves grounded context from the L2 data fabric, and proposes actions that are authorized, bounded, and executed only through L4 orchestration. L5 never writes to the hard-real-time control path. It advises; L1 and the autonomous hardware failsafe retain authority.
Three copilots compose L5. The Plasma Copilot designs shots and scenarios, optimizes operating points, and advises on disruption avoidance for the breeder and on plug/mirror operating points for the burner. The Engineering Copilot reasons over subsystems, triages anomalies, and plans maintenance. The Operations Copilot generates runbooks and procedures and supports shift handoff. All three share one substrate: retrieval-augmented generation over L2, context assembled from live twin state, agentic tool-use with bounded actions, and guardrails enforced at L4.
Design and simulation framing
Both Kronos machines are design-and-simulation studies today. The breeder (Hyperion) is a D-T spherical tokamak — Q_sci 3.076, 85.0 MW fusion power, 9.66 MA plasma current, 16.84 T peak field, negative triangularity delta -0.30. The burner (Aegis / MetroVolt) is a D-3He tandem-mirror generator with a 26.49 T plug and 17 T throat and 5.44% neutron fraction. Breeder construction begins Q2 2027 with first-of-a-kind first tritium targeted ~2030. The copilots are being built and validated against the twin now; none of their advice implies a hardware net-gain claim before FOAK.
- Inputs: L3 forecasts, L2 retrieval, live twin state vector
- Reasoning: LLM planner + tools, RAG grounding, uncertainty tags
- Outputs: proposed actions, ranked scenarios, cited explanations
- Authority: L4 authorizes; L1 executes; human approves side-effects
The remainder of this category documents each copilot, the RAG and context-construction machinery that grounds them, the bounded-action and authorization model that keeps them safe, the evaluation harnesses that measure them, and the human-oversight contract that keeps a qualified operator in the loop.