We propose a new procedure for inference on the parameters of the endogenous variables in a linear IV model that is robust to identification strength and is nonparametric with respect to the reduced form. The procedure adapts and generalizes the integrated conditional moment (ICM) test of Bierens (1982). The ICM procedure is first combined with the Anderson-Rubin approach that fixes the value of the parameter. This yields a first statistic that jointly tests whether the parameter value and the model are correct. We then develop a conditional test, that conditions upon an ICM-type statistic that informs on identification strength. In simulations, we find that our test has a well-controlled size irrespective of identification strength and is competitive with existing procedures.