Ole's original implementation used 'middle', which I switched away from.
After a great deal of testing it turns out that 'middle' does in fact
provide the most consistent results, so we're switching back to it.