{"discovery":{"accuracy":"(TP + TN) / |eligible universe|","aggregation":"micro-average from summed TP/FP/FN/TN counts","empty_set_convention":"precision is 1 when prediction and truth are both empty and 0 when only the prediction is empty; recall is 1 when truth is empty; F1 is 0 when precision plus recall is 0","f1":"2 * precision * recall / (precision + recall)","precision":"|prediction \u2229 truth| / |prediction|","recall":"|prediction \u2229 truth| / |truth|"},"end_to_end":{"all_true_consumers_protected":"every true affected consumer is present in the discovered set (recall == 1); extra consumers remain false positives and can still block readiness","correct_target_and_correct_patch":"exact patch-target selection and successful repair of every true target","ready_for_human_review":"successful composed row with exact consumers, exact targets, and accepted repairs"},"repair":{"attempts":"total repository attempts per trial; mean and median over every trial, including failures","first_attempt_acceptance":"accepted_attempt == 1 / all scored repair trials","within_budget_acceptance":"successful trials / all scored repair trials"},"schema_version":1,"statistics":{"confidence_level":0.95,"missing_observations":"report coverage and never impute values","paired_unit":"scenario_id plus repetition","proportion_interval":"two-sided Wilson score interval"}}