Early Recall Estimation and Stopping in Technology-Assisted Reviews via Ranker-Sampler Fusion

Technology-Assisted Review (TAR) workflows iteratively prioritize documents for expert screening to achieve high recall (e.g., 0.90 - 0.95) under limited time and budget. A persistent challenge is deciding when to stop: true recall is unknown during live screening unless all documents are judged, while additional sampling to estimate recall can be prohibitively costly at low prevalence. This paper develops a step-by-step statistical framework to estimate recall early by fusing (i) information from a ranking model and (ii) a small probability sample of documents. We present two complementary estimators: a Bayesian prevalence estimator that treats the ranker as an informative prior updated by a random sample, and a model-assisted estimator (generalized regression estimator with a Horvitz-Thompson residual correction) that combines calibrated ranker probabilities with design-based sampling adjustments. We then derive practical stopping rules based on conservative (lower-bound) recall estimates, and outline a future evaluation plan on established constrained high-recall benchmarks.