This paper develops an incentive compatible policy to control agricultural pollution, where the government knows the ranges of technology types and risk attitudes but not their distributions across farmers. The policy creates incentives for farmers to participate in the program, but includes constraints to ensure both self-selection of the appropriate policy, and self-compliance with the policy selected. Unknown risk attitudes are accommodated through stochastic efficiency rules. The model is applied empirically to estimate policies to limit nitrate contamination from New York agriculture. The estimated cost of such a program is not large compared to past commodity policies. Payments could be reduced if soils information is used to assign policies. Self-compliance is possible and does not impose a large cost on the government. If the policy were designed under risk neutrality, payments would be substantially below the incentive needed for participation by a risk averse farmer.