> ## Documentation Index
> Fetch the complete documentation index at: https://docs.latitude.so/llms.txt
> Use this file to discover all available pages before exploring further.

# Evaluation Triggers

> Configure which traces an evaluation runs against and when

# Evaluation Triggers

Every evaluation has a **trigger configuration** that determines which traces it evaluates. Triggers control monitoring scope without changing the evaluation's detection strategy.

## How Triggers Work

When a trace completes, Latitude checks it against each active evaluation's trigger configuration:

1. **Filter**: Does the trace match the evaluation's filter criteria?
2. **Sampling**: Should this matching trace be evaluated?
3. **Turn**: Is this the right turn in the session?

If a trace passes all checks, the evaluation runs. Otherwise, it skips that trace.

Triggers use the same **shared filter system** as trace views and saved searches.

## Trigger Fields

### Filter

Select which traces the evaluation monitors using any combination of shared filters:

* **Status**: Errors or successful traces
* **Models**: Specific models
* **Providers**: Specific providers
* **Tags**: Specific tags
* **Cost**: Above or below a cost threshold
* **Duration**: Above or below a duration threshold
* **Custom metadata**: Any `metadata.*` fields your application sends

An empty filter means "match all traces."

### Sampling

The percentage of matching traces that the evaluation runs against, from 0 to 100. Sampling controls cost and processing time while preserving coverage.

* Setting sampling to `0` effectively pauses the evaluation.
* New evaluations generated from issues default to `10%` sampling.

### Turn

Controls which trace or turn the evaluation runs on:

* **`every`**: Run on every completed trace (the default)
* **`first`**: Run only on the first trace or turn in a session
* **`last`**: Run only on the last trace or turn in a session

Use this when an evaluation only makes sense at the start or end of a conversation.

## Trigger Examples

**Monitor all production traces for jailbreak attempts:**

* Filter: metadata `environment` = "production"
* Sampling: 100%
* Turn: every

**Spot-check expensive traces for quality:**

* Filter: cost > \$0.50
* Sampling: 25%
* Turn: every

**Evaluate only the last turn of each session:**

* Filter: empty
* Sampling: 10%
* Turn: last

## Triggers, Search, and Annotations

Triggers scope automated monitoring. Search and annotations cover human review: use search to inspect relevant traces, then add annotations when you need human judgment for alignment or issue discovery.

[Flaggers](../annotations/flaggers) provide automatic signal for a fixed list of common categories.

## Next Steps

* [Alignment](./alignment): How human annotations calibrate evaluations
* [Evaluations Overview](./overview): How evaluations work
* [Annotations](../annotations/overview): The human side of the feedback loop
* [Search](../search/overview): Build cohorts of traces to review