# What are real-time guardrails?

Real-time guardrails check every prompt, response and tool call while the request is in flight, and block, rewrite or hold it before anything reaches the user or a tool. They prevent, where logs only report.

## Real time, as opposed to what

- **Offline evaluation** scores a model on a test set before release. It says how the model tends to behave.
- **Monitoring** reads logs after the fact. It says what happened.
- **Real-time guardrails** sit in the request path. They decide what is allowed to happen.

The three do different jobs: see [guardrails vs evals](/glossary/guardrails-vs-evals). Only the last one stops a leak before the data is out.

## Where they sit

1. **Before the model:** the user's prompt and any retrieved or fetched content.
2. **After the model:** the response, before the user sees it.
3. **Before a tool:** each tool call an agent makes, with its arguments.

## What makes it hard

A check in the request path has to be fast and right at once. Slow, and the product feels broken: see [guardrail latency](/glossary/guardrail-latency). Wrong, and it blocks honest users: see [false positives](/glossary/guardrail-false-positives). This is why real-time checks run on small purpose-trained [guard models](/glossary/guard-models) and not on a large general model asked to judge.

## With White Circle

[Control](/products/control) is real-time guardrails: it sits between your users, your model and its tools, checks each step against your policies, and blocks, rewrites or flags what breaks them. A small model gives verdicts in real time, and a larger one confirms them and is used to improve the small one.

