One of their hypotheses is that the test doesn't work because students can start a test with mental model A, get halfway through and realise it should actually be mental model B, but don't bother to go back to previous answers because the test is not used for assessment purposes.

Thus an otherwise excellent student could appear "inconsistent" in the predictive test, reducing its usefulness.

The second paper is good reading and its authors are admirable models of scientific honesty.