evaluation-quality
Scan evaluation data for n-gram overlap, canaries, answer patterns, hashes, and optional embedding similarity.
Generate controlled synthetic annotator disagreement before real reviewer data is collected.
