# Evals

> Tests that measure whether an AI system performs well on representative tasks or failure cases.

- Category: Evaluation & Safety
- Stage: Established
- Also called: Evaluations, Model tests
- Canonical: https://www.joelcsaji.com/ai-dictionary/evals

## Plain English

Evals are test suites for AI behavior.

## Example

A company tracks whether its support agent gives correct refund guidance across 200 realistic tickets.

## Why it matters

AI features drift and fail differently than normal software, so teams need behavioral tests.

## Related

- [Judge Model](https://www.joelcsaji.com/ai-dictionary/judge-model)
- [Golden Dataset](https://www.joelcsaji.com/ai-dictionary/golden-dataset)
- [Trace Grading](https://www.joelcsaji.com/ai-dictionary/trace-grading)

## Mentioned in

- [iOS 27, iPadOS 27, macOS 27, watchOS 27 and tvOS 27](https://www.joelcsaji.com/blog/ios-27-ipados-27-macos-27-watchos-27-tvos-27)
- [Fable 5, Export Controls, and GLM 5.2](https://www.joelcsaji.com/blog/fable-5-export-controls-and-glm-5-2)

## Sources

- [Anthropic: Building Effective Agents](https://www.anthropic.com/engineering/building-effective-agents)
- [Stanford HAI: 2026 AI Index](https://hai.stanford.edu/ai-index/2026-ai-index-report)