Abstract This article is an expansion on and a continuation of the ideas presented in Statman (Statman, S. 1992. Convergent tests and the divergent student. System, 20, 2) and at the Academic Committee for Research into Language Testing in 1993. It was then argued that the kind of convergent tests used frequently in EFL programmes are biased against the divergent student and suggestions for alternative modes of testing were put forward. It was believed then that the divergent student was the exception. Subsequent research indicates that varied performance is the norm. This paper seeks to explain the reasons for the many divergent answers; it suggests that the different interpretations of the text as it passes from original writer through test-maker and test-taker to test-grader with the inevitable distortion of the text by the questions leads to multiple interpretations. The different test-constructor/test-taker/test-grader interpretations are explained by reference to Spolsky's “turtles” model and illustrated by reference to student responses to tests administered both in class and in final examinations. Other factors include the difficulty of creating items which relate to only one level of comprehension, the tendency of students to be distracted by irrelevant details, student test-taking styles and the pressure of time. If, indeed, varied performance is more common than has been hitherto realized (or admitted), can comprehension be measured in any reliable fashion? If so, how? The paper comes to the conclusion that a portfolio of student work, compiled and graded after discussion of student “errors” with the student may be the fairest indicator of student reading ability. Some suggestions for alternate modes of test-construction for the mass screening procedures are also offered.