Glossary¶
Quick-reference alphabetical listing of key terms. Click any term for the full wiki article.
| Term | Definition | Page |
|---|---|---|
| Agent | LLM embedded in perceive-reason-act loop with tool access | agent-architecture |
| Attention sink | Initial tokens that stabilise transformer attention patterns | context-management |
| ATHEANA | HRA method identifying error-forcing contexts for commission errors | atheana |
| Automation bias | Over-reliance on AI recommendations without independent verification | automation-bias |
| Calibration | Degree to which expressed confidence matches actual accuracy | calibration |
| Capability gradient | 8-level prototyping path (L0-L7) for nuclear AI systems | capability-gradient |
| Cognitive load | Operator mental processing demands during agent supervision | cognitive-load |
| Common-cause failure | Failure of multiple barriers from shared cause | common-cause-failure |
| Compaction | Context compression by deleting tokens (zero hallucination risk) | context-management |
| Confidence-without-correctness | Property that hallucinated output reads identically to correct output | hallucination |
| Context divergence | Agents developing different situational pictures in shared room | context-divergence |
| Context rot | Performance degradation as input length increases | context-rot |
| Context window | Fixed-size token buffer processed in single inference pass | context-windows |
| DID / Defense-in-depth | Multiple independent safety barriers principle | defense-in-depth |
| Delivery modes | How agent outputs reach human participants (broadcast, flow gate) | delivery-modes |
| DPO | Direct Preference Optimisation — alignment without reward model | post-training-alignment |
| Epistemic independence | Decorrelated reasoning errors across agents | epistemic-independence |
| Error-forcing context | Conditions making wrong action appear correct (ATHEANA) | atheana |
| Fine-tuning | Adapting pre-trained model (SFT, LoRA, QLoRA) | fine-tuning |
| Flow gate | Delivery mode queuing agent outputs for sequential release | delivery-modes |
| Governance gate | Human approval required before safety-significant AI actions | human-authority |
| Hallucination | Fluent, confident, factually wrong LLM output | hallucination |
| Heartbeat | Scheduled periodic agent invocation without user trigger | agent-architecture |
| HRA | Human Reliability Analysis — quantifying human error probability | human-reliability-analysis |
| HSI | Human-System Interface — control room display architecture | hsi-architecture |
| IDHEAS-ECA | Macrocognitive HRA method (NUREG-2199) | idheas-eca |
| KG | Knowledge graph — structured entity-relationship database | knowledge-graphs |
| Lost in the middle | LLMs underuse information in centre of long contexts | context-windows |
| MCP | Model Context Protocol — model-agnostic tool connectivity standard | mcp |
| Model heterogeneity | Using different foundation models for independence | model-heterogeneity |
| Monoculture collapse | Correlated failure when all agents share one base model | monoculture-collapse |
| Non-determinism | Stochastic LLM output — same input may yield different outputs | non-determinism |
| Output vacuity / Slop | Superficially competent but substantively empty AI output | output-vacuity |
| Pattern 0 | Single-agent role simulation (baseline, no true multi-agent) | multi-agent-patterns |
| Pattern 7 | Personal assistant with heartbeat — persistent proactive monitoring | multi-agent-patterns |
| Pattern 9 | Shared room — concurrent agents, enforced isolation, human peer | multi-agent-patterns |
| PIF | Performance-influencing factor (IDHEAS-ECA terminology) | performance-shaping-factors |
| Prompt injection | Adversarial text in retrieved content altering agent behaviour | prompt-injection |
| PSA | Probabilistic Safety Assessment | psa-integration |
| PSF | Performance shaping factor (SPAR-H terminology) | performance-shaping-factors |
| Quantization | Model compression reducing weight precision (4-bit, 8-bit) | quantization |
| RAG | Retrieval-Augmented Generation | retrieval-augmented-generation |
| ReAct | Thought-action-observation agent loop pattern | agent-architecture |
| Reasoning models | LLMs with explicit chain-of-thought before final answer | reasoning-models |
| RLHF | Reinforcement Learning from Human Feedback | post-training-alignment |
| Room pause | Operator control halting all agent output delivery | human-authority |
| SA | Situation Awareness — perception, comprehension, projection | situation-awareness |
| SA Bridge | Agent role translating multi-agent output for operators | situation-awareness |
| Slop | See output vacuity | output-vacuity |
| Soul prompt | Persistent system prompt defining agent identity across sessions | prompt-engineering |
| SPAR-H | PSF-multiplier HRA method (NUREG/CR-6883) | spar-h |
| Structured output | Constrained generation (JSON mode, grammar-based) | structured-output |
| Sycophancy | LLM tendency to agree with user rather than maintain assessment | sycophancy |
| Tokenization | Sub-word segmentation of text into fixed vocabulary units | tokenization |
| Tool calling | Mechanism for agents to invoke external functions | tool-calling |
| Transformer | Neural network architecture using self-attention | transformer-architecture |
| Trust calibration | Matching operator trust to actual AI reliability | trust-calibration |