Stockfish's Curve Ball

During chess camp last week in a session ambitiously called "advanced rook endgames", I had the students choose a position from a stack of cards that were face down. These cards have 24 endgame diagrams, some of which specify the side to move. One of those they chose, I had regarded as reasonably simple. I set it up in the Fritz GUI with Stockfish 16 loaded as the main engine, and we (the students with my guidance) played it against Stockfish.

The engine defied Dvoretsky, who wrote, "Black's rook must remain passive, staying on the eighth rank" (Dvoretsky's Endgame Manual. 5th ed. [2020], 161). I explained to the students what I recalled of Dvoretsky's explanation, and drew the arrows visible in the diagram.

White to move

From my lead, the students chose the first move well.

1.Rb7 Rd8 2.Rh7

Only after struggling against the engine's surprising response did we come to understand that 2.Rg7+ is better.

2...Rd6!

3.f7 is not possible. We tried 3.Rg7+ Kf8 4.Rh7 Kg8 and realized something else was needed.
3.Re7 Kf8 confused us, although it should not have.
After 3.Rb7 Rd8, we felt that our progress was still missing something.

After several trials and error, we found the correct idea.

3.Rg7+ Kf8 4.Ra7 Rd8 5.Rh7

Black to move

Now Black must allow either the skewer or f7+ followed by the skewer.

Playing several Lucena positions against Stockfish over the past few years, I've learned that the engine often diverges from the line given in the textbook. Often this complicates the solution and the human must solve a problem.

These curve balls are why computer training is valuable.