Does anyone know of a formal from-first-principles justification for Occam's Razor (assigning prior probabilities in inverse proportion to the length of the model in universal description language)?

Maybe Kevin T. Kelly's work will fit your bill? Also see the discussion on LW.