Experienced rheumatologists differ widely in their assessments of rheumatoid arthritis even after extensive efforts to improve agreement by discussion and consenus. The use of computer feedback to provide an analysis of clinicians' judgment policies in a highly structured investigation has been shown to improve agreement, but this may not apply in normal clinical practice. Here the successful convergence of clinical agreements by three rheumatologists using computer assisted feedback over several months in a National Health Service outpatient department is reported. In the three months without feedback their pooled agreement for assessing the severity of rheumatoid arthritis was r2=0.62. During the three months in which feedback was provided agreement rose to r2=0.92. The principal component of all three judgment policies at the end of the feedback period was 'articular index'.