The system improves through corrections. When users mark a false positive or report a missed rep, the model adjusts. Over time, accuracy converges toward user expectations. Better to be honest than enthusiastic— undercounting builds more trust than overcounting.