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

# Build risk profiles

> Turn evaluation scores into a weighted risk assessment.

Risk profiles turn raw evaluation scores into one risk assessment per model. A risk category is a named risk dimension backed by evaluators scoped to datasets, and a profile is a weighted set of categories tuned for a use case. Scoring a profile against a model produces an overall risk score and risk level.

## Before you start

You need:

* At least one [risk category](/guard/app/risk-categories) to weigh.
* Evaluation results for that category's evaluators, produced by running them against datasets, usually through an [examination](/guard/app/examinations).

A category's score comes from stored evaluation results, so a profile built on categories that have never been evaluated produces nothing to score.

## Create a risk profile

<Steps>
  <Step title="Create New Profile">
    On the **Risk Profiles** page, click **Create New Profile**. To change an existing profile, click its card.
  </Step>

  <Step title="Describe the profile">
    In the **Profile** panel, set the **Name**, **Description**, **Use Case**, **Domain**, and **Method**. **Domain** offers government, healthcare, finance, and education. The **Active** toggle controls whether the profile is available for scoring.
  </Step>

  <Step title="Add categories and weights">
    Drag categories from **Categories** onto the **Profile Canvas**. Each card shows that category's aggregation method and evaluator count, and takes a **Weight** as a percentage. **Auto-Balance** distributes the weights evenly, and **Table View** shows the same categories as a grid.
  </Step>

  <Step title="Save Changes">
    Click **Save Changes**.
  </Step>
</Steps>

On the **Risk Profiles** page, profiles appear as cards showing the use case, domain, category count, a weight-distribution bar, tags, and the date modified. Profiles SeekrGuard provides carry a **System** badge. Each card's menu offers **Edit** and **Duplicate**. **Delete** appears only on profiles you created.

Above the cards, search by name, narrow the list with **Filter**, and filter by **Visibility**.

## Understand how a risk score is calculated

When a profile is scored against a model, SeekrGuard works bottom-up:

1. **Normalize each evaluator.** Each evaluation's score is scaled to a 0 to 1 range, `(score - min) / (max - min)`. For evaluators whose [polarity](/guard/glossary#polarity) is not positive, the value is inverted, so that higher means better at this stage.
2. **Score each category.** A category combines its evaluator and dataset pairings using its aggregation method.
3. **Aggregate to an overall score.** The profile's method combines the category scores by weight, then the result is inverted so that a higher score means more risk.

A category with no evaluation results is left out of the calculation rather than counted as zero. A model assessed on two of a profile's five categories is scored on those two, so scores are comparable only between models with the same coverage.

The overall score appears on a 0 to 100 scale and maps to a risk level.

| Risk score | Risk level |
| ---------- | ---------- |
| ≤ 40       | low        |
| ≤ 60       | medium     |
| ≤ 80       | high       |
| > 80       | critical   |

<Note>
  On the risk pages a higher score means more risk. The composite on a [model card](/guard/app/models) measures the same evaluations in reverse, where higher is better and 80 or above reads strong.
</Note>

## Read model risk scores

Per-model risk scores appear in the **Model Risk Assessment** section of the [Home page](/guard/app/navigate), where you select the profile to score against, and on each [model card](/guard/app/models), where you pick a profile in its scorecard. The model card breaks the score down by category, dataset, and evaluator under its **Risk Performance**, **Compare**, and **Explore** tabs.

<Note>
  Risk scores come from stored evaluation results. If a profile's categories have no evaluations behind them, there is nothing to score. Run the evaluators against their datasets first, usually through an [examination](/guard/app/examinations).
</Note>