A team has automated acceptance of high-confidence extractions. With a small review team, they want to keep measuring error rates, catch new kinds of errors, and spend reviewer time where it matters most. Which two practices should they adopt? (Select two.)

Answer: A and C — Review a stratified random sample of high-confidence extractions, stratified by document type / Route low-confidence fields and documents with contradictory values to reviewers first

Stratified random sampling of high-confidence output measures ongoing error rates and surfaces novel patterns across every segment (A), and routing low-confidence or contradictory cases first puts limited capacity where errors are likely (C). (B) undersamples rare types. (D) can't catch errors the model is confident about.