> 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.

# Decision Space

The Decision Space is where predictions and explanations can be evaluated, governed, and acted on.

In Umnai, a model can return both a [prediction](./explanations/local-explanations/predictions-and-attributions#prediction) and the [explanation](./explanations/local-explanations/overview) behind it. The Decision Space provides a way to apply user-defined logic and workflows to that output, so predictions can be scrutinised, adjusted, escalated, or approved in real time.

This turns a fit-to-data prediction into a fit-for-purpose decision.

## Identify, assess, resolve

Decision workflows can follow an Identify, Assess, Resolve, or IAR, methodology.

| Step     | Purpose                                                                                                                                        |
| -------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
| Identify | Use explanations to detect issues or sub-optimal characteristics in a prediction.                                                              |
| Assess   | Evaluate the impact, severity, and risk of the issue, including the consequences of resolving it in different ways or not resolving it at all. |
| Resolve  | Convert the prediction into a fit-for-purpose decision by recommending, ranking, prioritising, and executing an action.                        |

## Identify

The identify step uses explanation data to test for issues.

For example, a workflow might use control swaps to detect bias, activations to detect low data coverage, or attributions to identify which features and interactions contributed most to a prediction.

The goal is to identify whether the prediction has any characteristics that require further attention before it is used as a decision.

## Assess

The assess step evaluates what the identified issue means.

This can include:

| Assessment           | Example                                                                                             |
| -------------------- | --------------------------------------------------------------------------------------------------- |
| Impact               | Would the issue change the outcome?                                                                 |
| Severity             | How large is the effect?                                                                            |
| Risk                 | What is the risk of accepting the prediction as is?                                                 |
| Resolution trade-off | What happens if the system corrects the issue automatically, escalates it, or leaves it unresolved? |

Assessment helps determine whether the issue should be ignored, corrected, escalated, or handled by another process.

## Resolve

The resolve step turns the prediction into a decision.

Resolution can include:

| Resolution action                   | Description                                             |
| ----------------------------------- | ------------------------------------------------------- |
| Recommend actions                   | Suggest possible ways to handle the issue.              |
| Rank and prioritise recommendations | Decide which response is most appropriate.              |
| Execute automatically               | Apply the selected response without human intervention. |
| De-automate to a person             | Route the case to a human for review or final decision. |

This gives people oversight and control over how model outputs are used, while still allowing automation where appropriate.

## Example: identifying gender bias in loan approval

A decision workflow can use a control swap to identify and address potential `Gender` bias in a loan approval model.

One possible workflow:

1. Run the original query and run a control swap where `Gender` is changed.
2. Compare the prediction results. If the outcome does not change, continue to the final output step.
3. Compare the total attribution sum. If the difference is less than `10%`, change the outcome to `Loan Approved` and continue to the final output step.
4. Create a dashboard showing the control swap results, filtered to the top five differences, and request human input on the decision.
5. Output the prediction and save the audit log.

## Example: improving outcomes in loan approval

A decision workflow can also inspect whether a negative outcome is close enough to a threshold to deserve further review.

One possible workflow:

1. If the prediction is `Loan Not Approved`, measure the distance from the total attribution sum to the threshold.
2. If the distance is greater than `0.05`, continue to the final output step.
3. Identify any features whose values are close to activated partition boundaries. If no such features are present, continue to the final output step.
4. Create a dashboard showing the top five attributions together with dependency plots for features whose values are close to activated partition boundaries, and request human input on the decision.
5. Output the prediction and save the audit log.