The core is SPICE based (long ago) with LT enhancements
and extensions. For the supported device models I expect
its accuracy is the same, and controlled by the same
sort of tolerance variables, as other SPICEs. These settings
are the first thing to look at, along with solution method,
if you are looking for accuracy. For example, a TRAP
method when you are looking for noise qualities in a transient
simulation, is a bad idea because you tend to get a sawtooth
of TRTOL amplitude even on quiet nodes.
What to beware, is the accuracy of their models for complex
parts. These are macromodeled and LT warns you about
the possibility of detail-inaccurate results although they
seem to have done a reasonable job from the few ones
I have messed with.