The trait-test assumption of intelligence testing, according to which the test measures directly an innate trait, is contrasted with the conception that the test directly measures knowledge and that an innate trait could at best be measured only indirectly, the knowledge-test assumption. Two theories of knowledge accumulation are suggested, the linear learning theory, which assumes that knowledge accumulates as a linear function, and educational learning theory that postulates a positively accelerating function. A discussion follows of various phenomena related to the nonconstancy of the intelligence score, that have been largely overlooked apparently because the prevalent trait-test assumption does not account for them. It is shown that these phenomena fit well the predictions of a knowledge-test assumption combined with educational learning theory and that the nonconstancy is by itself rather predictable. This nonconstancy is then not an artifact of the “fallibility” of the measuring scale, but a fact that requires a revision in our dealing with and reporting of intelligence scores and in practices of prediction and experimental design.