Select two. Your extraction system scores 97% field-level accuracy on last quarter's validation set, and leadership wants to stop human review of every extraction the model marks as high confidence. Which two steps should come before you reduce human review?

Answer: A and C — Break accuracy down by document type and by field to confirm that performance holds in every segment, not only in the overall average. / Set up ongoing stratified random sampling of high-confidence extractions to measure their error rate and catch new error patterns.

An aggregate figure like 97% can hide poor performance on particular document types or fields, so accuracy has to be checked by segment before automating (a). Once high-confidence output skips review, stratified random sampling keeps measuring its error rate and catches new error patterns (c). (b) assumes confidence is calibrated and still doesn't show per-segment accuracy. (d) is the aggregate-metric trap itself.