Just as counter example, all the assignments for my OS class at top-20 CS university were graded using a set of a few hundred tests that were hidden from us. Fo

Just as counter example, all the assignments for my OS class at top-20 CS university were graded using a set of a few hundred tests that were hidden from us. For example, the first assignment was to implement a thread scheduler and over the years the course had developed a few hundred tests on which our grade was based. We were allowed to submit our solution three times, and we would be told how many tests we passed, but not why we failed any specific tests or what those tests were, with the highest of our submissions being our grade. The incentive of this process was two fold: first, it forced us to learn to write our own good tests and not rely on the TAs to spoon feed us exactly what we needed to do (in fact one student wrote a test that broke the reference solution and was added to the test suite for the next semester), and second it is completely intractable to expect a few TAs to scan through several thousand lines of code for each of a hundred students to discover how "close" to being correct their implementation is. Bugs are hard enough to find in code you wrote, let alone code written by an undergraduate just learning how to write systems code for the first time. In my experience TAing for both math and CS classes, finding bugs in students math assignments is significantly easier than finding bugs in code.