We basically learned about likely mistakes or discrepancies from our initial audits and fed that into our file selection algorithms. Once we sampled a few people who made a mistake, we were looking for everybody else who might have made the same mistake.
On June 8th, 2021. See this statement in context.