# Test a prompt with evaluations

**Category:** [Prompt Manager](https://help.totalctrl.app/hc/totalctrl/totalctrl-help-center/en-US/categories/prompts)
**Updated:** 2026-08-19

## How evals work

An evaluation runs a chosen prompt **version** against every row in a **dataset**. For each row, Prompt Manager compiles the prompt with the row's input variables, calls your account's LLM to produce an actual output, then uses an **LLM-as-judge** to score that output from 0 to 100 against the expected output. Scores are averaged into a run-level `avg_score`.
## Start a run

- Open a prompt and go to its **Evals** tab.
- Pick the **version** to test and the **dataset** to test against.
- Choose a **judge model** (a tenant or TotalCtrl-managed LLM profile), or leave it on the default. You can keep the LLM judge on or turn it off.
- Start the run.

## Small vs large datasets

Runs over datasets of **20 rows or fewer execute immediately**. Larger datasets are queued as pending and processed by a background job (roughly every five minutes, ten runs per tick), so the run finishes shortly after you start it.
## Read the results

The run detail page shows status, the average score, and a per-row breakdown of actual output, judge score, and rationale. You can **override any score manually** (0–100 with an optional rationale); the run average recalculates automatically. When a run completes, a `prompt.eval_run.completed` webhook fires with the average and counts.

---

## Related Articles

- [Prompts Manager Overview](https://help.totalctrl.app/en-US/articles/prompts-manager-overview)
- [Build a test dataset](https://help.totalctrl.app/en-US/articles/build-a-test-dataset)
- [Edit a prompt and save a version](https://help.totalctrl.app/en-US/articles/edit-a-prompt-and-save-a-version)
- [Promote a version with labels](https://help.totalctrl.app/en-US/articles/promote-a-version-with-labels)
- [GitHub Sync](https://help.totalctrl.app/en-US/articles/github-sync)

---
[← Back to TotalCtrl Help Center](https://help.totalctrl.app/en-US/)