I'm assuming the cardinal side is fairly standard. Metals are projected 6 at a time, and purification is used to detect when they match the target. One of the 5 unpurified metals is then used to replace the target atom, and the remaining 4 are output. This does waste 2/6ths of the metals, but I think the simplicity of initialization makes up for that.
The name refers to the gold detection method where it waits until every other metal is purified at the same time, which occurs after 60 tape loops (the LCM of 1,2,3,4,5).

This feels like a pretty efficient implementation. The only significant improvements I can see are replacing the pistons with normal arms and reducing the area swept by the initialization arm, but I wasn't able to find a good geometry for these. Cost is still the dominant term, so a cheaper but slower/bigger machine might beat this, but the hexarms are very convenient for pulling variable amounts of input and they naturally want to run at pseudo-period 3.