JUDGES AND RESEARCHERS COLLABORATED TO DETERMINE THE FEASIBILITY OF SENTENCING GUIDELINES AS A TOOL OF COURT POLICY TO REDUCE SHARPLY UNWARRANTED SENTENCING DISPARITY. THREE PRELIMINARY FEASIBILITY MODELS WERE DESIGNED TO DETERMINE HOW WELL SENTENCING DECISIONS COULD BE PREDICTED AND TO PROVIDE A VEHICLE FOR DISCUSSING THE ISSUES SURROUNDING THE UTILITY AND DEVELOPMENT OF GUIDELINES. NONE OF THE MODELS USED THE TWO-STAGE PROCESS BECAUSE THE INHERENT COMPLEXITY OF A BIFURCATED MODEL WOULD LIMIT ITS PRACTICAL APPLICATION. ALL THREE MODELS HAD TWO MAJOR CHARACTERISTICS IN COMMON: MODIFIED BURGESS WEIGHTING SYSTEMS AND DECISIONMAKING GRIDS. THE BURGESS APPROACH IGNORES INTERCORRELATIONS BETWEEN THE INDEPENDENT VARIABLES AND ASSIGNS ONE POINT TO EACH VARIABLE. THE HORIZONTAL AXIS OF THE DECISIONMAKING GRID IS DIVIDED INTO FOUR LEVELS OF SALIENT FACTORS AND THE VERTICAL AXIS INTO SIX LEVELS OF OFFENSE SERIOUSNESS. THE INTERIOR OF THE GRID CONSISTS OF 24 CELLS, EACH OF WHICH CONTAINS THE AMOUNT OF TIME A FEDERAL PRISONER IS TO BE INCARCERATED BEFORE BEING PAROLED. THE GRIDS PROVIDE A CONVENIENT PICTURE OF ACTUAL SENTENCING PRACTICES AND FACILITATES THE DEVELOPMENT OF SENTENCING POLICY. FURTHER, SUCH GRIDS MEET THE FOUR REQUIREMENTS OF ANY PREDICTIVE MODEL: SIMPLICITY, EFFICIENCY, RELIABILITY, AND VALIDITY. PREDICTOR VARIABLES WERE IDENTIFIED FOR INCLUSION ACCORDING TO THEIR RELATIONSHIP WITH THE 'IN/OUT' DECISION, THAT IS, THE DECISION TO INCARCERATE OR NOT TO INCARCERATE THE OFFENDER. MISSING DATA RULED OUT THE USE OF SCHOOL ATTENDANCE AND JOB STATUS INFORMATION. ALTHOUGH THE DECISION WAS MADE TO EXCLUDE EVEN PREDICTIVE ITEMS OF INFORMATION IF THEY POSED SIGNIFICANT ETHICAL OR LEGAL PROBLEMS IN INDIVIDUAL SENTENCING DECISIONS, ETHICAL CONSIDERATIONS DID NOT CEASE ONCE THE SET OF PREDICTOR VARIABLES WAS DEVELOPED. ALSO, BECAUSE THE JUDGES FELT THAT THE CATEGORIZATION OF AN OFFENSE AS SERIOUS WOULD USURP LEGISLATIVE AUTHORITY, A MATRIX SYSTEM WAS DEVELOPED WITH ONE GRID FOR EACH OF THE FELONY/MISDEMEANOR CLASSES. ONCE IMPLEMENTED, THE DEMONSTRATION MODEL ACCURATELY PREDICTED 90 PERCENT OF THE 'IN/OUT' DECISIONS IN ITS CONSTRUCTION SAMPLE. IN A VALIDATION SAMPLE, THE MODEL CORRECTLY PREDICTED 88 PERCENT, WITH 8 PERCENT FALLING OUTSIDE THE GUIDELINES IN TERMS OF LENGTH OF INCARCERATION. THE JUDGES CONCLUDED THAT SENTENCING GUIDELINES COULD BE DEVELOPED USING RELATIVELY SIMPLE TECHNIQUES AND THAT THEY COULD COLLABORATE WITH RESEARCHERS TO ESTABLISH EXPLICIT REPRESENTATIONS OF A COURT'S SENTENCING POLICY. TABULAR DATA, REFERENCES, AND IMPLEMENTATION INFORMATION ARE PROVIDED. (KBL)