Glossary category
Evaluation and reliability
The evidence and controls used to decide whether an AI system works well enough to trust.
What belongs here
Evals, benchmarks, verification, observability, failure analysis, and reliability measures.
What belongs elsewhere
Security controls whose primary purpose is authorization, privacy, or abuse prevention.
Adjacent categories
Start with
Factory areas
Definitions in this category
stable
A/B testing
A randomized controlled experiment that exposes comparable groups to different variants and compares a predefined outcome.
Ablation
An experiment that removes or changes one component while holding others as constant as practical to estimate that component's contribution.
Acceptance criteria
Explicit conditions an outcome must satisfy before it can be accepted, promoted, or declared complete.
Accuracy
The proportion of evaluated predictions that are correct under a specified labeling and decision rule.
Benchmark
A standardized set of tasks, data, metrics, and procedures used to compare systems under a defined evaluation regime.
Calibration
The degree to which stated probabilities or confidence levels correspond to observed frequencies of correctness.
Concept drift
A change over time in the relationship between inputs and the correct target or decision.
Confusion matrix
A table counting predicted classes against actual classes, including true and false positives and negatives.
Cross-validation
A resampling method that estimates generalization by repeatedly training and evaluating on different non-overlapping subsets of available data.
Data drift
A change over time in the distribution of system inputs or features.
Evaluation (eval)
A systematic measurement of model or system behavior against defined tasks, criteria, datasets, or operational outcomes.
F1 score
The harmonic mean of precision and recall.
False negative
An outcome incorrectly classified as absent when it is actually present.
False positive
An outcome incorrectly classified or flagged as present when it is actually absent.
Held-out set
Evaluation examples deliberately withheld from training, prompt tuning, workflow design, or agent feedback to reduce leakage and overfitting.
Nondeterminism
The property that identical-looking requests can produce different behavior because of sampling, concurrency, infrastructure, model updates, or hidden state.
Observability
The ability to infer a system's internal state and behavior from emitted traces, logs, metrics, events, and artifacts.
Precision
Among predicted-positive cases, the proportion that are truly positive.
Recall
Among truly positive cases, the proportion correctly identified as positive.
Reward hacking
Achieving a high measured reward through behavior that exploits the metric or evaluator without accomplishing the intended objective.
Specification gaming
Satisfying the literal specification or metric in a way that violates its intended purpose.
working
Abstention
An explicit outcome in which a model or evaluator declines to answer, act, or judge because evidence or authority is insufficient.
Golden set
A curated collection of examples with trusted expected outcomes used for regression testing or evaluation.
Grader
A component that scores, classifies, or judges an output or trajectory against a rubric or expected behavior.
Independent verification
Checking an outcome with evidence, components, context, or authorities meaningfully separated from the system that produced it.
LLM as judge
Using a language model to evaluate, compare, classify, or score outputs produced by models or agents.
Oracle
A mechanism that can determine the expected or acceptable result for a task, such as a compiler, formal specification, invariant, test suite, or reference implementation.
Sycophancy
A failure mode in which a model favors agreement with a user's stated belief or preference over an independently supported answer.
Trace
A captured sequence of model calls, tool calls, events, timings, state changes, and outputs from an execution.
Verification gate
A control point that blocks promotion until required evidence has been produced and validated.
Verification loop
A repeated execute, observe, compare, and correct cycle that withholds completion until an attempted result satisfies explicit evidence or acceptance criteria.
contested
Emergent behavior
A capability or behavior that appears qualitatively new at a larger scale or higher level of system interaction rather than as an obvious continuation of smaller-scale measurements.
Model drift
A broad operational term for model behavior or performance changing relative to an accepted baseline.