Ian, your post gave me a bit of deja vu.
Bruce, that sounds very similar to a bug report I gave you (back when Demeter was in alpha?) on Nov 3 2011. You might want to check your personal Email, or I could forward a copy to the group. Even though it's many versions ago this might be a different facet of the same bug. Relevant text:
"Take any of my "Ref" samples from the files I gave you and start a new file. (Alternatively, use any data you have.) Make a quick FS fit and let Artemis fill in the Path variables. (Although I see this on all types of fits, and when I write my own variables.) Make one fit with a "bad guess" for one parameter, like N*so2=1 for my FCC Pd, and for the next fit, make a proper guess of 12. (This is only to deliberately make fits with very different results; this may also work by varying the window, etc.) Copy the log for each result and set aside for a moment. Now, go to the GDS window and hit "Discard All." Go back to the History tabs and look at the fit parameters again. You should see the report from the last fit is unchanged, and the statistical parameters for the first fit are the same, but the guess parameters and everything below for the first fit are now overwritten by the values from the last fit."
-Jason