I've measured the relative error (for each value in each decade, the rest set to zero, and the "zero" resistance, 0.5 Ohm, subtracted) of my two 4-decade assemblies (built using MELF 0.1% resistors) on HP3458A. One unit (D0) uses selected resistors, the second (D1) - the rest. As an afterthought - it is possible to get a better relative error by measuring the resistors and arranging them in a certain way (placing the lowest error resistor as the first in the decade, and then go for the resistor giving the lowest error for the sum of the first two and so on, as the weight of the absolute error is diminishing for each next resistor). Next time I'll try that approach.
Cheers
Alex