So I've been using Rally for a few days on Zwift and the accuracy is great as long as I keep the power down the whole time. It's almost exactly 1:1 with a Quarq DZero being recorded concurrently and compared in DCRainmaker's analysis app. Seriously, at any given point on the mean-max graph, the power was within 1W even down to 30s duration. However I just rode a TTT in Zwift, which is a lot of ON/OFF and the overall average dipped significantly. Shorter intervals where power was steady were once again within 1W. Longer intervals would include several instant power dips with no corresponding cadence drop, and thus lower the average.
I'm wondering if anyone else has observed this behavior...