Evaluation Sandbox

Use the Evaluation Sandbox to test model behavior against representative inputs, scoring rubrics, and selected AI connections before you approve a workflow for wider use.

​

Open the sandbox

1. Open Organization → Sandbox, or visit /my-org/evals.
2. Start on Guided evaluation.
3. Choose the content type and test-input type.

​

Run a guided evaluation

1. Enter the System Prompt and required Test input.
2. Choose an existing rubric, or create scoring criteria in the rubric library.
3. Select Change AI connection or model and choose the model connection to test.
4. Use Advanced parameters when you need generation controls.
5. Review the test count and select Run tests.
6. Review the results, record observations, and save useful inputs as scenarios when the action becomes available.

​

Reuse evaluation assets

• Use Scenario library for reusable inputs.
• Use Conversation tests for multi-turn behavior.
• Use Red Teaming for adversarial testing.
• Use Connected tests for integrated workflows.
• Use Saved evaluations and Run history to revisit repeatable tests.
• Use Rubric library for reusable scoring criteria.

​

Evaluation checklist

• Test the normal case, incomplete inputs, and edge cases.
• Use synthetic or approved data.
• Record the model connection and rubric used.
• Treat failures as signals to refine instructions, validation, or permissions.
• Do not promote an evaluation result to production without the organization’s approval process.