My composition, white to play and draw.
I was tempted to add 'testposition' to the subject, but the position is more useful to make fun of engines than to test them.
The point is that after 1.Nxg5 hxg5 white is lost, because the black knight is only 13 steps away from b3 (good luck for your favourite engine to see it). But 1.Bf3 with the idea Nf2-e4-d2 (and if black gets too creative, timely Ne4 and g5, Bh5, counterattacking d6 and f6) holds the fort.
Of course, engines misevaluate resulting positions (thinking black should be content with draw after 1.Nxg5), but I find it striking that for humans it is easy to see the knight journey. I think that is because humans distinguish dynamical position (which attack is quicker) from statical (fortress-like) ones. Engines are good at dynamical ones, but in statical positions they are not so good. In my opinion it is about move ordering and reductions. In dynamic positions it is good to look at moves with good history first, but in statical position it is better to first try moves without history (to avoid getting stuck in local optima). Maybe it is not feasible to teach engines this distinction, maybe new-moves-first would hurt the search too much, and it would be better simply to improve static evaluation in closed positions.
Anyway, the position is not very natural, so no big deal.
