In this paper we describe a version of the
level method for solving a nondifferentiable
program where the problem data (feasible set,
functionvalues, and subgradients) are not known
explicitely and may be computed using an oracle
only up to a appropriately chosen accuracy
eps We show that the modified level method
produces a delta^{dom}-feasible,
delta^{opt}-optimal solution after a finite
number of oracle calls for positive
delta^{dom}, delta^{opt}.
Further, we describe how our general algorithm
can be applied to decomposition problems which
are generally nondifferentiable. In such an
application the assumption of eps-exact data
is natural since the inner program of the
decomposed problem in general is nonlinear and
hence solvable only approximately.