A forecasted defect count is modeled two ways: exponential decay by 12% per sprint or a linear reduction of 12 defects per sprint. Which evidence would favor the exponential model?