Essentially, we need the hashrate displayed by the miner's device to match the hashrate we calculate based on the difficulty of the shares (hashes) it submits.
Can't you accomplish that by making the shares small enough, so the miner has multiple shares per reward period? That way you should get a nice average instead of peaks.
A real-world example: during local testing with CPUMiner, 50% of the shares come in with a difficulty between 0.05 and 0.15 at a rate of 10 shares per minute. However, episodically, CPUMiner calculates shares with a difficulty of 10.0 and even 40.0 - and this completely skews the hashrate calculation over 10-minute intervals. I've tried trimming extreme values in the calculations, but I still can't achieve acceptable accuracy between the hashrate reported by CPUMiner in its logs and the hashrate we calculate. I'm currently experimenting with different normalization schemes, but haven't found a suitable solution yet. So I thought - maybe someone has a ready-made solution.