Nuclear AI Wiki

A living knowledge base on AI agent systems for nuclear power plant operations. Compiled from a six-report series by Michael Hildebrandt (IFE, 2026).

Glossary — Alphabetical quick-reference for all key terms.

Source Summaries

• summary-llm-foundations — Report 1: LLM architecture, agent systems, failure modes, knowledge grounding
• summary-multiagent — Report 2: 10-pattern taxonomy, coordination, epistemic independence
• summary-nuclear-regulatory — Report 3: NRC regulatory implications, safety analysis, control room design
• summary-scenarios — Report 4: Nine nuclear control room scenarios as worked examples
• summary-hra — Report 5: HRA method adaptation for AI-assisted operations
• summary-prototyping-guide — Report 6: 8-level capability gradient for building and testing

Core Concepts

LLM Fundamentals

• transformer-architecture — Self-attention, unified context processing, lost-in-the-middle effect
• tokenization — Sub-word tokens, implications for numerical reasoning
• context-windows — Fixed-size buffer, context rot, usable capacity
• statelessness — No memory between invocations
• non-determinism — Stochastic output, implications for V&V
• post-training-alignment — SFT, RLHF, Constitutional AI, DPO

Failure Modes

• hallucination — Confident fabrication, confidence-without-correctness problem
• output-vacuity — Slop: superficially competent but substantively empty output
• calibration — Overconfidence, degradation on hard questions
• sycophancy — Agreement bias from RLHF training
• automation-bias — Over-reliance on AI recommendations
• trust-calibration — Matching operator trust to AI reliability
• prompt-injection — Adversarial inputs altering agent behaviour
• context-rot — Performance degradation as input length increases

Knowledge & Grounding

• retrieval-augmented-generation — RAG architecture, failure points, context rot
• knowledge-graphs — Structured constraints, guardrails, nuclear procedures as graphs
• structured-output — JSON mode, grammar-constrained decoding
• fine-tuning — LoRA, QLoRA, when to fine-tune vs use RAG
• quantization — GPTQ, AWQ, GGUF, accuracy trade-offs
• reasoning-models — Extended thinking, visible reasoning chains

Agent Architecture

• agent-architecture — Perceive-reason-act loop, ReAct pattern, memory types
• tool-calling — Function calling, reliability concerns, MCP
• prompt-engineering — System prompts, soul prompts, sensitivity
• context-management — Compression, summarisation, context discipline

Multi-Agent Systems

• multi-agent-patterns — 10-pattern taxonomy (Patterns 0-9)
• epistemic-independence — Decorrelated reasoning errors
• monoculture-collapse — Common-cause failure from shared models
• model-heterogeneity — Using different foundation models
• delivery-modes — How agent outputs reach humans
• context-divergence — Agents developing different situational pictures
• degradation-and-recovery — Graceful degradation, fallback modes
• situation-awareness — SA Levels 1-3 mapped to LLM architecture
• cognitive-load — Operator supervision limits, attention management

Nuclear Operations

• defense-in-depth — Layered barriers, epistemic analogue for AI
• graded-autonomy-tiers — Three regulatory tiers (Tier 1-3)
• human-authority — Governance gates, room pause, flow gates
• alarm-prioritization — AI-assisted alarm management
• shift-handover — SA continuity across crew changes
• emergency-response — AI advisory during LOCA and emergencies
• procedure-ai-interaction — AI interaction with EOPs, AOPs, Tech Specs
• hsi-architecture — Display design for AI advisory
• common-cause-failure — Shared failure modes across AI systems

HRA & Safety Assessment

• human-reliability-analysis — HRA methods overview
• spar-h — PSF multiplier model
• idheas-eca — Macrocognitive function decomposition
• atheana — Error-forcing contexts
• performance-shaping-factors — Standard and AI-specific PSFs
• psa-integration — Incorporating AI into PSA event trees

Prototyping & Testing

• capability-gradient — 8-level path (Levels 0-7)
• build-vs-assess-gap — Buildability outpacing assessability
• evaluation-harness — Structured LLM evaluation
• local-deployment — Running models behind air-gap boundaries
• simulator-coupling — Connecting LLMs to RELAP5, FRAPCON
• operator-modelling — Simulating operator cognitive states

Comparisons

• rag-vs-fine-tuning — When to retrieve vs when to train
• single-agent-vs-multi-agent — When multi-agent is justified
• frontier-vs-local-models — Cloud API vs local deployment
• hra-methods-compared — SPAR-H vs IDHEAS-ECA vs ATHEANA

Scenarios

• scenario-shift-handover — SMR shift handover (Pattern 7)
• scenario-rcs-temperature — RCS temperature anomaly (Pattern 0 vs 9)
• scenario-concurrent-alarms — Concurrent alarms during transient (Pattern 9)
• scenario-safety-verification — Independent safety verification (Pattern 9 + diversity)
• scenario-loca — Loss of Coolant Accident (Pattern 9, 7 agents)
• scenario-fuel-monitoring — Long-term fuel integrity (Pattern 9 + KG)
• scenario-instrument-failure — Instrument channel failure (Pattern 9)
• scenario-multi-unit-smr — Multi-unit SMR monitoring (9 agents)
• scenario-compound-event — Compound event with ambiguous indications

Entities

Organizations

• nrc — U.S. Nuclear Regulatory Commission
• iaea — International Atomic Energy Agency
• epri — Electric Power Research Institute

Frameworks & Tools

• crewai — Role-based multi-agent framework
• autogen — Microsoft conversational multi-agent
• langgraph — Graph-based agent orchestration
• letta — Memory-focused agent framework
• mcp — Model Context Protocol
• ollama — Local LLM serving

Simulation

• relap5 — Thermal-hydraulic simulation
• frapcon — Fuel performance simulation

Reactor Types

• pwr — Pressurised Water Reactor
• smr — Small Modular Reactor

Standards & Regulations

• 10cfr50 — NRC nuclear power plant regulations
• nureg-0700 — HFE review guidelines
• nureg-0711 — HFE engineering process
• nureg-1792 — Good practices for HRA
• nureg-2199 — IDHEAS-ECA methodology