Criminal justice reforms are currently advocated by both liberal and conservative criminologists and penologists. Liberals advocate decriminalization of status and victimless offenses, diversion of lesser offenders from legal processing, due process for all, and deinstitutionalization of correctional programs. Conservatives call for a return to law-and-order oriented penal philosophies, such as just deserts sentencing, determinate sentences based upon prior history and current offense, incapacitation of chronic offenders, and deterrence. The focus of evaluative research should be to carefully test each innovation, instead of accepting it outright as an improvement over existing practices. Experimental analysis should therefore be applied, providing for assessment of process and outcome of all criminal justice reforms and innovative programs. Rigorous evaluation in the form of field experimentation has redeeming social value; however, many impediments to experimentation are represented by political and bureaucratic interests, professional versus scientific interests, situational problems, and legal considerations. Even when actual practice is not feasible, a model for field experimentation should take into account such criteria as control, representativeness, and naturalness. The latter criterion is exemplified by using the natural, ongoing processes of the community to reintegrate offenders rather than attemting to create new forms of community-based institutional control for the same purpose. A working field experimentation model should incorporate such elements as mutual research and project goals, a carefully defined target population, a formal theoretical statement of the problem, an intervention strategy, a research strategy, and an assessment of implications. The model should be administered by an experienced, research-oriented person, possibly a university-affiliated investigator (e.g., a doctor rather than a practitioner in the helping professions), who can help design the various components of the research model and convey to his staff the viewpoint that research must play a neutral role. A total of 61 references is appended.