Glossary
Key terms used throughout the AIRE Standards.
A
Active Learning Process where an AI system learns from human corrections (HITL feedback) to improve over time. Corrections are added to the golden dataset and used to retrain the model.
AIRE AI Reliability Engineering. The discipline of building and operating reliable AI agents, inspired by Site Reliability Engineering (SRE).
Agent An autonomous AI system that uses an LLM for reasoning and can take actions (call APIs, query databases, etc.) on behalf of users.
C
Chain of Thought (CoT) The LLM's step-by-step reasoning process before arriving at a final answer. CoT logging captures this reasoning for debugging and observability.
Circuit Breaker Fault tolerance pattern that stops calling a degraded service after repeated failures, preventing cascading failures. Has three states: Closed (normal), Open (failing fast), Half-Open (testing recovery).
Cognitive Reliability The practice of ensuring AI agent outputs are accurate, consistent, and trustworthy through validation, self-reflection, and drift detection.
Confidence Calibration Ensuring an agent's reported confidence score correlates with actual accuracy. A well-calibrated agent with 90% confidence should be correct 90% of the time.
D
Drift Detection Monitoring for changes in input distributions, output characteristics, or model performance over time. Types include input drift, output drift, and model drift.
Durable Execution Workflow execution pattern where state is persisted so workflows can resume from the last checkpoint after failures, rather than restarting from scratch.
E
EDD (Eval-Driven Development) AI equivalent of Test-Driven Development (TDD). Development methodology where golden datasets and evals guide development, and deployments are blocked if evals fail.
Ephemeral Context Context (prompts, conversation history) that is processed but never persisted to logs or storage, used for highly sensitive data like medical records.
F
Fallback Path Alternative strategy when primary system fails. Example: GPT-4 → GPT-3.5 → Rule-based system → Human.
Feedback Loop System where production failures and HITL corrections are collected, added to the golden dataset, and used to retrain models, creating continuous improvement.
G
Golden Dataset Curated collection of inputs with labeled expected outputs, used as a regression suite for evaluating model performance before deployment. Typically 100+ examples.
Groundedness Percentage of claims in an agent's output that are supported by retrieved context or known facts. Target: >95%.
Guardrails Deterministic hard limits that constrain agent behavior regardless of LLM reasoning. Types: input guardrails (PII redaction, prompt injection detection), output guardrails (sensitive data leakage prevention), action guardrails (rate limits, monetary limits).
H
Hallucination When an LLM generates factually incorrect or fabricated information presented as fact.
Hallucination Rate Percentage of agent outputs containing factual errors. Target: <0.1%.
HITL (Human-in-the-Loop) Design pattern where humans review or approve agent actions before execution, typically for high-stakes decisions or low-confidence outputs.
I
Idempotency Token Unique identifier that ensures retrying a failed operation doesn't duplicate side effects (e.g., double-charging a customer).
J
JIT (Just-in-Time) Privilege Access Security pattern where agents are granted minimum necessary privileges scoped to specific actions with automatic expiration (typically <5 minutes).
Jailbreak Attack where adversarial input manipulates the LLM to bypass intended constraints or reveal system prompts.
M
MTTD (Mean Time to Detect) Average time from when an issue occurs to when it's detected. Target: <5 minutes.
MTTR (Mean Time to Recovery) Average time from failure detection to full recovery. Target: <5 minutes.
O
Offline Evals Pre-deployment evaluations run on a golden dataset in CI/CD to catch regressions before they reach production. Cheap and reproducible.
Online Evals Post-deployment evaluations run on production traffic to detect drift and real-world failures. More expensive but catches unknown issues.
P
Prompt Injection Attack where user input contains instructions that override the system prompt, causing the agent to behave incorrectly.
R
Resumability Ability of a workflow to resume from the last checkpoint after a failure, rather than restarting from scratch.
Rollback Reverting to a previous version of a model or system after detecting performance degradation in a new deployment.
S
Scoped Token API key or access token limited to specific actions and resources (e.g., "refundOrder:12345"), reducing blast radius of security breaches.
Self-Reflection Pattern where an LLM critiques its own output before finalizing, useful for high-stakes decisions. Example: Agent generates answer → critiques it → revises answer.
Staged Rollout Deployment strategy where new models are rolled out gradually (5% → 50% → 100% traffic) with monitoring at each stage.
Structured Output
LLM output constrained to a specific format (JSON schema, enum) to enable deterministic validation. Example: {"answer": "...", "confidence": 0.9}.
T
Tool
External function or API that an agent can call to perform actions (e.g., getWeather(), sendEmail(), queryDatabase()).
This glossary is part of the AI Reliability Engineering (AIRE) Standards. Licensed under the AIRE Standards License v1.0.