Estimate Recall When You Have No Complete Ground Truth

Your deduplication pipeline for 5 million records reports 96 percent precision on a reviewed sample of predicted matches. Management asks: "what share of the real duplicates do we find?" You only have the output of the pipeline and a team of reviewers who can spend about two weeks.

1. Explain why random pair sampling is the wrong way to estimate recall.
2. Design a record-level sampling plan, give the estimator and say how you would report uncertainty.
3. Explain how blocking recall and classifier recall combine, and how you would measure each separately.
4. A colleague proposes capture-recapture with two independent blocking families. When does it work and when does it mislead?