fastsimcoal non-parametric bootstrap: model.par vs model_maxL.par

[Theodor Liljeros]

Dear Laurent,

I’m having trouble understanding how to correctly perform the non-parametric bootstrap in fastsimcoal.
Following the manual, I use the model.par file of my best model and run:
./fsc25 -i model.par -n1 -q -j -s0 -d -b9
to generate non-parametric bootstrap SFS.

What confuses me is this:
Some parameter values in model.par are much higher than the MLEs reported in the corresponding .bestlhoods file from my original model.
When I then run fastsimcoal on the bootstrap SFS (using the same .tpl and .est as in the original analysis), the parameter estimates from the bootstrap replicates tend to cluster around the values in model.par, rather than around the original MLEs. This makes it look as though the .par values used for -b are effectively anchoring the bootstrap, which feels more like a parametric bootstrap.
So my questions are:

For the recommended “non-parametric bootstrap” procedure using -b, should I really be using the original model.par file as described in the manual, even if its parameters differ strongly from the MLEs?
Or is it more appropriate in practice to base this .par file on the model_maxL.par (i.e. the MLEs of the best model)?
Thank you very much for your time and clarification.

Best regards,
Theo