Edit-log lag stats are censored: use Kaplan-Meier plus a period-by-age hazard table
Taking the median of (correction year - submission year) over records that have been corrected biases the lag downward. Recent cohorts can only show short lags, and records never corrected drop out. Use Kaplan-Meier instead: entry at the submission year, event at the first correction, censoring at the current year. Report cumulative incidence at fixed horizons by cohort. To tell steady discovery from one clean-up campaign, tabulate the hazard (events per 1,000 record-years at risk) by calendar period x age band. A campaign appears as a spike in one period, while steady discovery gives a stable rate. Exclude age 0 if fresh records get edited during review. On the OEIS this showed no plateau: about 0.15 percentage points a year at 10–25 years.
Follow-up: the same censoring trap applies to raw 'share ever corrected' by category. OEIS 'hard' vs 'easy' sequences looked 2.5x apart in raw shares (3.6% vs 1.4% for 2010+ entries). Run per group with KM on 1996-2014 entries, they were nearly equal at 5 years (2.7% vs 2.6%) and only modestly apart at 10 (4.1% vs 3.4%). The categories differed in age mix and exposure. Compare groups at fixed horizons, not raw shares.