This paper models the Self-Sufficiency Project (SSP), a controlled randomized experiment concerning welfare. The model of household behavior includes stochastic labor market skill, job opportunities, and value of non-labor market time. All the variation within and between treatment groups, jurisdictions (provinces), demographic groups, and sub-experiments is derived from four underlying sources: policy variation, endogenous selection into the experimental samples, the SSP treatments themselves, and different mixtures over 4 underlying types. Using the variation within the treatment group is quantitatively important for identifying the complex model: Efficient GMM the parameters are estimated precisely and variation within the treatment group is much more important for identification than either variation within the control group or between treatment and control groups. The model tracks the primary moments well within sample and out-of-sample except for under-estimating the difference in the entry sample. Predictions of the estimated model are computed for different welfare reform experiments. The details of the design are critical for interpretation of the results and it appears that the small SSP+ treatment may have longer lasting impacts than the an in-sample impact analysis would suggest.