Capability Gradient

The capability gradient is an 8-level prototyping path for developing and evaluating nuclear AI advisory systems. Each level adds capability incrementally, and critically, nothing is discarded — capabilities accumulate. Levels are experiments that produce evidence, not technologies to be deployed.

The Eight Levels

L0 — Frontier Evaluation: Evaluate frontier models on nuclear-domain questions using the evaluation-harness. No deployment, no infrastructure. Produces baseline capability data. Achievable in a day.

L1 — Local Deployment: Run models locally behind air-gap (see local-deployment). Tests data sovereignty, prompt transparency, and capability at reduced model scale.

L2 — RAG + KG: Add retrieval-augmented-generation and knowledge-graphs for domain grounding. Tests retrieval quality and constraint enforcement.

L3 — Tool Calling: Enable tool-calling to external systems (historian, databases). Tests integration reliability and error compounding.

L4 — Simulator Coupling: Connect to thermal-hydraulic codes via simulator-coupling. Tests physics-grounded projection capability.

L5 — Persistent Agent: Add persistence, memory, and background monitoring (see operator-modelling). Tests long-running operation and state management.

L6 — Multi-Agent: Deploy multi-agent-patterns. Tests coordination, context-divergence, and epistemic-independence.

L7 — Human-in-Loop: Full operator interaction in simulator setting. Tests human-authority, trust-calibration, automation-bias, and hsi-architecture.

Running Example

A single scenario — RCS temperature anomaly — threads through all levels, growing in sophistication from a simple question-answer evaluation (L0) to a full multi-agent advisory supporting an operator in a simulator (L7).

Evidence Production

Each level produces evidence addressing specific gaps identified in Reports 1-5. L0 establishes baseline capability. L2 tests whether grounding resolves hallucination. L4 tests whether simulation coupling enables meaningful projection. L6 tests whether multi-agent coordination produces situation-awareness benefit. L7 tests whether operators actually benefit or suffer from automation-bias.

Timeline

L0 is achievable in a day. The full gradient from L0 through L7 represents 6-12 months of systematic development. The gradient approach addresses the build-vs-assess-gap by ensuring that assessment evidence is produced at each level rather than deferring all evaluation to the end.