Chapter 3 of 6
Confirm the following:
- The platform inventory names the exact nonproduction Foundry resource, project, policy-assistant agent, approved version, candidate version, and stable-endpoint selector. The platform owner confirms that its stable endpoint selects the approved version. (The model-governance, governed-agent, Agent 365 access-boundary, and MCP security controls establish these prerequisites.)
- The approved and candidate versions are different immutable versions of the policy-assistant agent.
- The operator and project managed identity have Foundry User on the exact project.
- A network-isolated project has the approved delegated evaluation subnet.
- The quality owner has confirmed same-day regional and evaluator support.
- The tool owner has approved tool-call evaluators for the Function Tool path.
- The cost owner has approved judge-model and evaluation consumption.
- The golden set contains synthetic content only. Each JSONL row stays below 2 MB, and the batch stays within 100,000 rows.
Implementation files#
| Type | File | Consumer |
|---|---|---|
| Runtime | artifacts/eval/evaluation-spec.json | The evaluation runner, release gate, preflight scripts, and Session 12 validators |
| Runtime | artifacts/eval/data/golden-v1.jsonl | The evaluation runner and release gate |
| Deployment | artifacts/eval/thresholds.yaml | The release gate, preflight scripts, and Session 12 validators |
| Deployment | artifacts/release/release-policy.json | The release owner, release gate, preflight scripts, and Session 12 promotion workflow |
| Record | artifacts/operations/disable-and-restore.md | The release owner and incident operator |