If you believe that you have correctly guessed what the optimal randomized
algorithm is, you can verify your claimed competitiveness by testing it
against all oblivious deterministic adversaries.  If your algorithm is
defined using the distributional model, and it is stable, i.e., for a each
work function you give a specific distribution, such an adversary can be
defined by what it does at each work function, and there are never more
than two choices, since there are ony two tasks the adversary can request.
and not that many work functions.  I would suggest defining a potential
at each work function.

One way to prove the lower bound is to find an optimal oblivious randomized
adversary.
