Then, thanks to both of you, I think it is safe to say that the case is solved :

the hermicity of the hamiltonian is 'ensured' by Kwant by recalling the same function (and taking the complex conjugate) for i--->j and for j--->i, but in the case of a RNG in the function, this yields to different values.

But using a table of random numbers generated with a hash function 'kwant.digest.uniform()' works fine !

Thank you very much for your help !

Alexandre