Evaluation design decisions for a large-scale child care provider training program are discussed as a function of three program characteristics: decentralization, local autonomy, and multiple delivery strategies. The fact that delivery strategies were not assigned, but evolved as a logical result of local autonomy, presented particular difficulties to evaluation. Rather than using a sampling plan, we gathered data for all training courses and trainees. While decentralized process data were gathered, only centralized outcome data were collected. Consideration of the relative merits of the approaches taken indicated that the decisions made wereappropriateforaprogram with such diverse implementation.