We seem to get better results when we scale bridge IPs down to multiples of 1 instead of 8 and only mark bridges as blocked if they differ in the 'bad' direction by at least one standard deviation from the mean.