Evaluate a grader

Freeze the human labels before running the grader. Use one narrow criterion with a written rubric and structured output. The grader must be able to return unknown or request human review.

Keep human labels outside the judge workspace. Use python3 -m helpers.evals.cli run-isolated-judge to create a fresh child workspace containing only the named examples and current rubric. Run every revised rubric in a different fresh child. A child must not receive human labels, prior verdicts, captured verdicts, or later rubric versions. Preserve the isolation record with the verdicts.

Use helpers.evals.judges.evaluate_alignment for the confusion table and disagreement set. Repeat at least one unchanged boundary example and use repeated_label_stability to measure scoring stability.

Inspect:

• every human and grader disagreement;
• label imbalance;
• an answer-order reversal when judging pairs;
• a verbosity trap where a longer answer is not the better answer;
• an ambiguous example that should produce unknown; and
• whether the generator and grader are actually independent contexts.

Revise one rubric criterion at a time and rerun only the working examples. Do not tune on the heldout judge set.

Call the classroom result an alignment check. A small agreeing sample is not completed calibration.