> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.umnai.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.umnai.com/_mcp/server.

# Local explanations

Use a local explanation to inspect the prediction for an individual record. Before querying, confirm that the intended deployment is running and that the input values match its schema.

## Before you begin

Have a **RUNNING** deployment and one record expressed in its input schema. Check numeric units, categorical labels and missing-value conventions. Use the record's actual values; the prefilled form values are examples, not a substitute for the observation you want to explain.

## Run a local explanation

1. Open **Explanation Lab > Local Explanations**.

2. Choose the model in **Deployed Model**. The example selects **Adult Income**.

3. Enter the feature values for the record you want to inspect.

4. Select **Execute Query**.

5. Open **Query Result** and review the summary and **Raw Query Result**.

Field labels follow the deployed schema. The example includes age, gender, relationship, and encoded labels such as **capital_45_gain**, **educational_45_num**, and **hours_45_per_45_week**. Match the fields shown in your own deployment.

## Use optional diagnostics

Use **SAW Mode**, where available, for an additional diagnostic view. Enable the switch, supply the target values requested by the form and execute the query. Review **SAW Output** together with the prediction and input record. Availability depends on the deployment; use the [REST API strengths and weaknesses guide](/guides/views/deployed-model-views/strengths-and-weaknesses) for further detail.

**Strength** indicates how strong the prediction appears; lower values can warrant review. **Uncertainty** reflects disagreement between estimators and is most informative when the model has multiple estimators. **Adjusted error** compares the prediction with the supplied target observation. Read these diagnostics together with the record and its attribution rather than treating any one value as a guarantee of reliability.

## Interpret the result

| **Example result** | **Displayed value** |
| ------------------ | ------------------- |
| Prediction         | 0.0795957328        |
| Total Attribution  | -2.4479319078       |
| Model Intercept    | -1.0337139777       |

These values describe the example query. Results change with the model and inputs. The prediction applies the model's output transformation, so it is not the simple sum of the displayed total attribution and intercept.

Check the inputs before interpreting the result. Special values such as **\_\_MASK\_\_** and **\_\_INFREQUENT\_\_** follow the model's preprocessing conventions. Use them only when they are appropriate for the selected model and observation.

![Select a deployed model and enter the record to explain.](/_fern-files/_img/f066c7ea9517d00e1992fe78c2850ae35c3698b689c1154d3ced4dffb8c81fc3.webp)

![Query result summary and raw JSON for the example record.](/_fern-files/_img/e687491ec87da199eece32791a6b9dfee86d5649b913cdc29b6e77505a00806c.webp)

Read **Prediction** as the model's output for the submitted record. **Total Attribution** and **Model Intercept** describe the explanation before the output transformation. Use the returned activation and model type when relating them to the prediction; do not treat an attribution as a probability or percentage change.

Use **Raw Query Result** when you need the returned values and resource context rather than a rounded display. Keep the input record and deployed-model identity with any result you reuse.

## Next steps

Continue to [Feature and module attributions](/ui-guides/explanations/attributions).