Calibrate a grader before using it as a deployment gate

A grader score is another model output until it is compared with a trusted label. Agreement alone is insufficient when one error type matters more: a grader that misses unsafe failures can look accurate on a mostly passing set.

Define the measurable control

Write a narrow rubric from observed failure modes. Label examples independently, resolve disagreements and hold out a calibration set the grader did not shape. Report the confusion matrix, sensitivity, specificity and performance by slice. Set the deployment threshold before measuring.

The AI reliability lab makes the decision reproducible with authored data:

from reliability_cases import grader_calibration_case

result = grader_calibration_case()
assert result["held_out_labels"] == 100
assert round(result["sensitivity"], 3) == 0.808
assert round(result["specificity"], 3) == 0.938
assert result["gate_allowed"] is False

On one hundred held-out labels, the fixture’s grader finds 42 failures and misses 10. Sensitivity is about 80.8%, below its 85% minimum, even though specificity is about 93.8%. The grader is therefore barred from being the release gate.

Challenge the result

Repeatedly tuning on the holdout converts it into training data. Create a new holdout after material rubric or prompt changes. Preserve human review for ambiguous and high-impact cases, and track whether production distributions drift from calibration data.

Keep a reviewable evidence pack:

• rubric and labelled calibration set
• confusion matrix with slice metrics
• pre-agreed gate thresholds
• disagreement review and recalibration trigger

This work aligns with the evaluation, security, cloud operations, reliability and FinOps sequence in the FDE for Professionals course. The linked course describes the learning pathway; this article’s numbers are synthetic and do not report a model, client, audit or production result.

Practice task

Label fifty pass/fail cases, hide the labels, run a deterministic mock grader and compute both false-positive and false-negative consequences.

Continue learning

This article is part of the Advanced AI reliability and assurance sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Create a model-change regression release process.
• Continue with Design a production evaluation sampling strategy.

Reference: OpenAI evaluation guidance.