Inference on experimental data when the quality of randomization is unknown