Terms
Every term in this hub.
This hub collects every evaluation term in the Prefactor glossary in one place, from the fundamentals to the sharp edges. If you are building your first eval suite, start with the core measurement concepts: ground truth evaluation, golden datasets, and the difference between automated and human evaluation. If you are already scoring runs in production, the metrics group covers groundedness, answer faithfulness, context precision and recall, and the quality scores that roll individual checks into a per-agent signal.
Two clusters deserve special attention. LLM-as-judge techniques, where one model scores another's output, made subjective quality measurable at scale; the entries here cover the method, its known biases, and the rubrics that keep it honest. And regression detection is where evaluation earns its keep: catching the moment a model update, prompt change, or data shift makes a previously good agent worse, before users notice.
The terms are grouped below by how teams meet them in practice: measurement fundamentals first, then metrics, judging methods, datasets and benchmarks, testing practices, and finally the failure modes evaluation exists to catch, from hallucination to reward hacking. Each entry links to a full definition, and most connect onward to the guides that show the applied side.
Evaluation vocabulary overlaps with observability on purpose: you cannot score what you never captured. Where a term straddles both, it lives in the cluster teams search it from, and its related-terms list bridges the gap. If a term you expect is missing, the A-Z list on the main glossary page covers every entry across all clusters.
Measurement fundamentals
Automated Evaluation Automated evaluation uses programmatic checks, model-based judges, or statistical metrics to assess agent performance at scale. Canary Evaluation Canary evaluation routes a small fraction of production agent traffic to a new agent version and monitors quality, error rate, and policy compliance in real time before full rollout. Ground Truth Ground truth is a verified, authoritative dataset of correct answers or expected outcomes used to evaluate an AI agent's accuracy. Ground Truth Evaluation Ground truth evaluation is the assessment of an AI agent's outputs against a known-correct reference dataset to measure factual accuracy, task completion, and output quality. Human Evaluation Human evaluation is the process of having people assess an AI agent's outputs for quality, accuracy, helpfulness, and safety.
Metrics
Answer Faithfulness Answer faithfulness is an evaluation metric that measures whether an AI agent's response is grounded in and consistent with the source context it was given, rather than hallucinated or contradicted by the context. Confidence Score A confidence score is a numeric value representing how certain a model or agent is about a particular output or decision. Confidence Score (Agent) A confidence score is a numerical estimate of how certain an AI agent is about its output or the correctness of a decision. Context Precision Context precision is a RAG evaluation metric that measures whether the retrieved documents an agent was given actually contain the information needed to answer the query, as opposed to returning many irrelevant chunks that increase noise and token cost. Context Recall Context recall is a RAG evaluation metric that measures whether all the information needed to answer a query was present in the retrieved context. Groundedness Groundedness measures whether an AI system's output is supported by its source material. Quality Gate (AI) A quality gate is an automated evaluation checkpoint that an AI agent's output must pass before it is delivered to an end user or triggers a downstream action.
Judging methods
Agent Grading Rubric An agent grading rubric is a structured scoring guide that defines the criteria and weighting used to evaluate agent outputs, covering dimensions such as accuracy, completeness, format compliance, safety, and tone. Human Preference Annotation Human preference annotation is the process of collecting human judgements, typically choosing between two model outputs or rating quality on a scale, to measure subjective dimensions of agent quality that automated metrics cannot capture, such as tone, helpfulness, and trustworthiness. LLM-as-Judge LLM-as-judge is an evaluation technique where a separate language model, typically a capable model like GPT-4 or Claude, is used to score the outputs of an agent being evaluated.
Datasets & benchmarks
Testing practices
A/B Testing (Agent) Agent A/B testing compares two versions of an agent, differing in model, prompt, tools, or configuration, by routing a portion of traffic to each and measuring performance, quality, and user satisfaction differences. Evaluation Harness (Agent) An agent evaluation harness is the test infrastructure that automatically runs an agent against a suite of benchmark tasks, captures outputs, scores them against defined criteria, and generates performance reports. Evaluation Pipeline An evaluation pipeline is an automated workflow that benchmarks agent quality, accuracy, safety, and policy compliance before and after deployment. Prompt Regression Testing Prompt regression testing is the practice of re-running a fixed set of evaluation cases against an AI agent every time a prompt, model, or tool configuration changes, to catch quality or safety regressions before they reach users. Regression Detection (Agent) Regression detection for AI agents is the monitoring of quality and performance metrics over time to identify when a previously acceptable agent has begun producing worse outputs: due to model drift, data distribution shifts, prompt degradation, or changes in connected tools. Regression Testing (Agent) Agent regression testing verifies that changes to a model, prompt, tool, or policy have not degraded the agent's performance on previously passing test cases. Safety Testing Safety testing is the process of deliberately probing an AI agent with adversarial, edge-case, and boundary-condition inputs to verify that safety controls hold. Stress Testing (Agent) Agent stress testing evaluates how an AI agent performs under extreme conditions: high request volumes, very long inputs, rapid tool call sequences, or resource constraints.
Failure modes
Configuration Drift Configuration drift occurs when an AI agent's deployed state diverges from its approved, governed configuration over time, through manual changes, dependency updates, or environment differences. Drift Detection (Agent) Agent drift detection monitors whether an agent's real-world behavior has diverged from its expected behavior baseline. Hallucination A hallucination occurs when an AI model generates information that sounds plausible but is factually incorrect or fabricated. Hallucination Rate Hallucination rate is the percentage of an AI system's outputs that contain fabricated or unsupported claims, measured over a defined sample. Model Drift Model drift is the gradual change in a model's behavior or output quality over time, often caused by shifts in input data, fine-tuning updates, or upstream model provider changes.
More agent evaluation terms
Agent Grounding Agent grounding is the practice of anchoring an agent's responses and decisions in verified, authoritative data rather than relying solely on the model's parametric knowledge. Agent Marketplace An agent marketplace is a platform where pre-built agents, tools, or plugins can be discovered, evaluated, and deployed. Agentic RAG Agentic RAG combines retrieval-augmented generation with autonomous agent capabilities, allowing the agent to decide what to retrieve, from which sources, and how to synthesize results. AI Impact Assessment An AI impact assessment is a structured evaluation conducted before deploying an AI system to identify potential risks, benefits, and mitigations. AI Maturity Model An AI maturity model is a staged framework that describes the progression from ad hoc, ungoverned AI usage to optimised, fully governed AI operations. AI Procurement AI procurement is the process of evaluating, selecting, and onboarding external AI models, tools, or services with appropriate governance due diligence. Algorithmic Impact Assessment An algorithmic impact assessment is a structured evaluation of how an AI system may affect individuals, communities, and society, covering risks such as discrimination, privacy loss, economic displacement, and erosion of trust. Answer Relevance Answer relevance is an evaluation metric that measures how directly an AI agent's response addresses the user's actual question or task. Bias Detection Bias detection is the process of identifying systematic unfairness in an AI model's outputs across protected groups or sensitive attributes. Catastrophic Forgetting Catastrophic forgetting is when a neural network loses previously learned knowledge after being trained on new data. Constitutional AI Constitutional AI is an alignment approach where a model is trained to follow a set of principles or rules. Data Poisoning Data poisoning is an attack where an adversary manipulates training or retrieval data to influence a model's outputs. Deceptive Alignment Deceptive alignment is a hypothetical failure mode where an AI agent behaves correctly during training and evaluation but pursues a different objective when deployed in production, having learned to recognise when it is being monitored. Embedding An embedding is a numerical vector representation of text, images, or other data that captures semantic meaning. Fine-Tuning Fine-tuning is the process of further training a pre-trained model on a domain-specific dataset to improve its performance on particular tasks. Fundamental Rights Impact Assessment A fundamental rights impact assessment evaluates how a high-risk AI system may affect individuals' fundamental rights, including privacy, non-discrimination, freedom of expression, and due process, before and during deployment. Knowledge Base A knowledge base is a structured or unstructured collection of information that an AI agent can search and retrieve from to inform its responses. MLOps Engineer An MLOps engineer builds and operates the infrastructure for training, evaluating, deploying, and monitoring machine learning models and AI agents at scale. Model Card A model card is a standardised documentation format for an AI model that discloses its intended use, performance benchmarks, known limitations, training data, evaluation methodology, and ethical considerations. Model Distillation Model distillation is the process of training a smaller, faster model to replicate the behavior of a larger model. Model Governance Model governance is the set of controls that manage the selection, approval, deployment, monitoring, and retirement of AI models used within an organisation. Model Validation Model validation is the independent verification that an AI model performs as expected across accuracy, fairness, robustness, and safety benchmarks. Multi-Model Strategy A multi-model strategy is the practice of using different language models for different tasks, agents, or risk levels within an organisation. RAG (Retrieval-Augmented Generation) An architecture pattern where an AI agent retrieves relevant information from external knowledge sources before generating a response. Reflection Reflection is an agent capability where the model evaluates its own output or reasoning before returning a final answer. Reflection Agent A reflection agent is an AI agent that critiques and revises its own outputs before returning a final result. Retrieval Agent A retrieval agent is an agent specialised in searching, filtering, and returning information from knowledge bases, databases, or APIs. Retrieval Poisoning Retrieval poisoning is an attack that corrupts the external data sources an agent retrieves from, such as knowledge bases, vector stores, or document repositories, to manipulate the agent's behavior through its RAG pipeline. Risk Scoring Risk scoring assigns a numeric or categorical value to an AI system based on its inherent risk factors, control effectiveness, and operational context. RLHF (Reinforcement Learning from Human Feedback) RLHF is a training technique that refines a model's behavior using human preference judgments. Shadow Agent An AI agent deployed within an organization without the knowledge or approval of security or IT teams. SOC 2 (Service Organization Control 2) An auditing framework that evaluates how a service organization manages customer data based on five Trust Service Criteria: security, availability, processing integrity, confidentiality, and privacy. SOC 2 Type II SOC 2 Type II is an independent audit report that evaluates whether a service organisation's controls around security, availability, processing integrity, confidentiality, and privacy were operating effectively over a defined observation period, typically six to twelve months. SOX (AI Controls) SOX, the Sarbanes-Oxley Act, requires US public companies to maintain internal controls over financial reporting. Specification Gaming Specification gaming occurs when an AI agent satisfies the letter of a task specification while violating its intent, finding loopholes in the way success was defined rather than doing what was actually wanted. Structured Output (AI) Structured output is the generation of model responses in a defined machine-readable format, such as JSON, XML, or a typed schema, rather than free-form text. Technical Documentation (AI) AI technical documentation is the detailed record of an AI system's design, development, testing, and deployment that regulators may require.
Questions
Common questions.
What is agent evaluation?
Agent evaluation is how teams find out whether an AI agent actually did its job: whether the response was correct, grounded, on budget, and delivered the outcome the task needed.
What is LLM-as-judge?
LLM-as-judge is an evaluation technique where one model scores another model's output, making subjective quality measurable at scale. The entries in this hub cover the method, its known biases, and the rubrics that keep it honest.
How do teams catch agent regressions?
Through regression detection: re-running evaluations when a model, prompt, or data source changes, to catch the moment a previously good agent gets worse, before users notice.