This paper demonstrates that an electronic tutoring program can collect data that enables a far more reliable assessment of students' skills than a standard examination. Socratic electronic homework tutor, CyberTutor can integrate effectively instruction and assessment. CyberTutor assessment has about 62 times less variance due to random test error than a three-hour final examination, and 53 times less variance than 12 weekly tests. This improvement results from the combination of three factors. First, the typical student interacts with CyberTutor 16 times longer than the three-hour final exam and 10 times more than 12 weekly tests together, diminishing the effect of lucky guesses and careless errors. Secondly, by considering requests for hints, solutions, and, importantly, the number of wrong answers and the time the student takes to complete each part of the problem, CyberTutor can make a better determination of the student's skill on a problem. Lastly, if a problem is too difficult for the student as presented, CyberTutor's hints and responses to wrong answers systematically adapt the difficulty of the problem to the student's skill level. (Contains 10 references.) (Author/AEF)