This plot shows the noise with the box on, but no granite. We're still pretty far off from the Guralp data sheet.
I implemented software rotation in the huddle subtraction as Valera suggested and it works much better. The two plots below show the before and after. So far this is just 2 deg. of rotation around the z-axis. I'm assuming that aligning the seismometers vertically via bubble level is good enough for the z-axis, but I haven't calibrated the bubble yet.
The residual slope is now suspiciously smooth. I somehow suspect that our readout electronics can still be responsible. We need to hook up a 9V battery to the input terminals to check it out. Its a little steeper than 1/f and I thought that we had exonerated the Guralp breakout box in the past, but now I'm not so sure. I'll let Jenne comment on that.
I also noticed that we have not yet divided by sqrt(2) to account for the fact that we are subtracting 2 seismometers. In principle, an unbiased estimate of the single seismometer noise will be lower by sqrt(2) than the green curve.