If you're having developers do the same task twice, without reusing the results from the first round, you're got a problem at a different level.

But, seriously, the "yesterday's weather" approach to estimating suffers from several problems. For one, the context may be different. Dave may have taken N days to write the last widget because Susan (who knows a lot about widgets) was sitting nearby. Susan has since moved on. Is it now going to take Dave N days to write another widget? Maybe yes, but probably no.

For another, Dave may have spent a lot of time climbing up the learning curve the first time. What took N days then make take N/2 now.