Chapter 4 of 6
Decisions and stop conditions#
Set required values through the approved delivery change path. Keep project endpoints, resource IDs, subscriptions, tokens, prompts, responses, tool payloads, and personal data out of the repository.
Set the gate policy#
Before the baseline run:
- Record the approved Foundry evaluation definition ID in
evaluation-spec.json. The runner retrieves that definition and starts a run; it does not create another definition. - The quality owner approves the relevance floor.
- The tool owner approves the tool-call accuracy and success floors.
- The safety owner approves safety floors of
1.00. - Every blocking metric has a maximum error count of
0.
After the run, set each active quality or tool threshold no lower than its owner-approved floor and no higher than the approved baseline result. Conflicting bounds stop the gate. Fix the approved version or change the owner floor through the delivery path. Do not weaken a floor to fit the baseline.
The threshold policy, release policy, and approved aggregate must name the same baseline run ID.
Keep the target and evaluator boundary fixed#
Run the complete gate in East US 2 while protected material remains blocking there. If that path is unavailable, stop. Move the gate or have the safety and release owners revise the policy before another run.
Keep the gate disabled if the evaluated path adds Azure AI Search, Bing Grounding, Bing Custom Search, SharePoint Grounding, Code Interpreter, Fabric Data Agent, or Web Search. The tool owner must approve compatibility again.
The gate returns PASS only when every blocking metric is present, meets its threshold, and has zero evaluator errors. Preview evaluators cannot become the sole blocking control.
An exception may cover an eligible final-answer quality failure. It must name the metric and business reason, compensating control, owner, exception authority, and expiry. It cannot override a safety or tool-process failure or enable automatic promotion.
Stop when a version is mutable or implicit, the project or region is wrong, a limited-support tool enters the path, a preview evaluator becomes blocking, or a safety or tool-process failure is treated as overridable.