# Guardrails vs evals

Guardrails check live traffic and act on each request, while evals measure a system before release on a fixed set of test cases. They answer different questions, and a production system needs both.

## What each one is

An eval (evaluation) is a test run: a fixed set of inputs goes through the system, the outputs are scored, and the result is a number such as accuracy or the share of attacks that succeed. Automated [red teaming](/glossary/ai-red-teaming) is an eval whose inputs are attacks.

A [guardrail](/glossary/ai-guardrails) is a check in the request path. It judges each live input, output and tool call against a policy and acts: block, rewrite, flag.

| | Evals | Guardrails |
| --- | --- | --- |
| Runs on | A fixed test set, before each release | Live traffic, on every request |
| Result | A score for the system | A verdict and an action per request |
| Catches | Regressions and known weak spots | The single bad request or reply |
| Misses | Inputs outside the test set | What no policy describes |

## Which to pick

Both. Each covers the other's gap.

- Evals alone measure the failure rate and stop none of the failures.
- Guardrails alone stop what their policies describe, and you learn that a new model or prompt made things worse only once it is live.

They also feed each other: failures caught in production become eval cases, and an eval of the guardrails shows what they miss and what they block by mistake. A new policy can be tested the same way on past traffic ([policy backtesting](/glossary/policy-backtesting)).

## With White Circle

[Control](/products/control) enforces your policies on live traffic in real time. Red Team tests the system with simulated attacks, in one-off campaigns or continuously, with regressions tracked over time.

