It is also well known (I highly recommend to read the article by David Goldberg "What Every Computer Scientist Should Know About Floating-Point Arithmetic") that if we change an algorithm in a program with floating point arithmetic, the results of the program will also change. It should be noted that PARDISO uses dynamic parallelization technique. That means that the order of operations is determined dynamically at run time and as a consequence the order of operations might varies from run to run. That leads to the deviation in the results since round-off errors might be different.
The condition number is, in fact, the measure of sensitivity of the solution to round-off errors. There exists estimates for the admissible errors in solution of linear systems. These estimates are well known and can be found in books on linear system solvers (see for example the book G.W.Stewart Matrix Computations") and these estimates are based on a careful analysis of impact of round-off errors on the solution. So lets' take a simplest estimate which is the following
|| (exact solution) - (computed solution) || <= eps*( the condition number of the system) * || right hand side||
where eps is the relative precision for a given type of input data. For example in the case of single complex arithmetic operation eps is 1.1920929E-07, for double precision eps is about 1^(-16).So the larger condition number is, the larger deviations might be.
All the best