The necessary input to SGP4 is a TLE, and a time. It takes the difference between the epoch date of the TLE and the chosen time and plugs it into a gigantic pile of complicated maths, which generates position and velocity, or whatever other representation of the state you request, at the desired time. Because it's using polynomials of trig functions of polynomials with time as the only variable, some kinds of errors oscillate, but others grow quadratically or worse with time. You can ask SGP4 to plug in any time interval you want, but six months will produce gigantic errors.
What exactly are the contents of your "archived set of orbital ephemerides"? Is it just a bunch of TLEs for the same object at different epoch times, or do you actually have position and velocity to compare against? Also, which version of SGP4 are you talking about? What does this code you found look like? There are a number of very different versions, and I don't yet know which you might have found.