# False positives in guardrails

Guardrail false positives are harmless prompts, responses or tool calls that a guardrail blocks or flags by mistake. A stricter policy catches more violations and also stops more legitimate traffic, so the rate has to be measured and tuned.

## Precision and recall

**Precision** is the share of flags that are real violations. False positives lower it. **Recall** is the share of real violations that get flagged. Misses lower it. Raising one tends to lower the other.

Violations are rare, so a small error rate goes far. If 1 request in 1,000 breaks a policy and the check wrongly flags 1 harmless request in 100, it raises about ten false flags for every real violation.

## Why they matter

Each one is a user refused for a reasonable request, or an agent stopped mid-task. Users rephrase or leave. Teams stop reading the flags, then switch the policy off, and it protects nothing.

## How to reduce them

- **Backtest.** Run a changed policy over past traffic before it goes live, and read what it flags. See [policy backtesting](/glossary/policy-backtesting).
- **Scope.** Apply a [policy](/glossary/guardrail-policy) only where it belongs: one feature, one event type, one user segment.
- **Review.** [Flag](/glossary/guardrail-actions) uncertain cases for a person instead of blocking them. Their rulings show where the policy is wrong.

## With White Circle

[Improve](/products/improve) backtests a new policy against historical traffic and scopes policies per feature, use case, event type or user segment. Analyze has a human review queue for edge cases.

