Updated: The first version of this question was implicitly assuming that we know in advance roughly what the distribution of grades in a class will be (for instance, roughly the same as last year). This assumption wasn't actually essential to the question, and was bothering some people, so I've updated it to remove that assumption.
Inspired by this question, how difficult should an exam be? For instance, assume you've decided about what skills should denote a B- student, and that you're free to norm exam scores to course scores as needed. By adjusting the questions you put on an exam, you can adjust what score a student with B- skills will typically get. Where should you aim for?
Some answers I can think of, which some plusses and minuses:
Aim for the score traditionally correlated with that grade. For instance, in the US there's a custom that a B- corresponds to a grade in the 80-82 range, so you could try to write an exam where a score of 80-82 will typically reflect the skills of a B- student. (Actually, if it's a class where non-exam items will tend to increase grades, aim a bit lower.) The advantage is that, if the exam, there's "no curve", and students tend to feel like they've done well. A disadvantage is that if you accidentally make the exam too easy (say, a question has a short cut and doesn't reveal as much as you intended), your options are limited.
Aim a bit lower. Say, aim for an exam where a score of around 70 reflects a B-. More importantly, this gives a bit more separation at the top of the class (you have room for a question or two to distinguish the top students from the very good). On the other hand, students seem to feel like their grades are arbitrary, and they may feel more competitive (and less willing to work and study together) if they feel like their friends doing well will cut into their curve.
Aim much lower---say, put a B- around 50 and curve heavily. This makes the cut-offs between final grades much wider (the gap between a B- and a B might be something like 7 points instead of 3), and therefore a single small mistake has a smaller impact on a score. Students get quite discouraged by these, in my experience.
My question is whether there's any actual research suggesting that some of these are clearly good/bad for student learning, and whether there are ethical or practical issues I'm overlooking.