Label eight responses. Set the limits. Check the release.
Reference labels decide the release gate. This eight-case authored sample is not a model benchmark or a reliability estimate.
How the check works
Pass rate uses the authored reference labels. Daily cost uses mean request cost × traffic. P95 uses complete request times and nearest rank; with eight observations, it is the slowest request.
OpenAI’s evaluation guide calls for task-specific tests and human calibration. Our data, labels, and costs are invented. No model call; nothing is stored or sent.