Keetman vs Van Arkel: The Promotion That Changed Everything
Three decisions from a sharp final-round game, from a missed mating net to a rare rook underpromotion.
Reproducible analysis run complete
6 saved models on NVIDIA GeForce RTX 5080 · Stockfish 18 · 250,000 nodes per checked line.
Why this game?
A public thank-you to Amsterdam Chess Open
Jan Timmerman and Stichting Amsterdam Chess Open welcomed an original review from the 2026 event. This page thanks the organizers by naming both players, linking the official tournament page and complete public broadcast, and turning the dramatic Keetman-Van Arkel finish into an interactive lesson.
- White
- Maaike Keetman
- Black
- Rogier van Arkel
- Event
- Amsterdam Chess Open, Group A, round 6
- Result
- 0-1
Interactive comparison
Choice one: keep the mating net intact
White to move after 34...Ba4. Keetman played 35.d6 and stayed clearly better, but Stockfish finds a forced mate with 35.Nbc5. The six models instead divide between two losing rook captures.
k2n1b2/p1R2pp1/Np5p/3PPp1P/b7/PN6/1q3PP1/3Q2K1 w - - 5 35Model candidates
Choose a viewpoint
V6 Active: Rxa7+
Measured GPU choice. The check looks forcing, but Black takes the rook and the winning attack disappears.
Model top three: Rxa7+ · Rxf7 · Na5
Stockfish 18 safety gate
CompleteStockfish finds mate in nine with 35.Nbc5. The played 35.d6 remains strongly winning at about +5.25, while both measured model captures lose the attack and fail the gate.
- Engine move
- Nbc5
- Depth
- 22
- Node budget
- 250,263
- Mate
- M9
Verification line
35.Nbc5 Bd7 36.Rxd7 Nb7 37.Rxb7 Qxe5 38.Nc7+ Qxc7 39.Rxc7 Bxc5 40.d6 Bxf2+ 41.Kxf2 Kb8 42.Qf3 a6 43.Qb7#
Position lesson
The advanced pawns look urgent, but the key is to keep every attacker. Nbc5 threatens the black king while protecting the rook and preserving the d-pawn. Capturing on a7 or f7 removes a vital rook from the attack and gives Black time to escape.
Interactive comparison
Choice two: promote before adding another attack
White to move after 36...Be7. Every model promotes with 37.d8=Q+. Keetman played 37.Qd5, kept an advantage, but made the finish harder.
Model candidates
Choose a viewpoint
V6 Active: d8=Q+
Measured GPU choice. Promotion with check keeps a decisive advantage and passes the engine gate.
Model top three: d8=Q+ · d8=R+ · Rxc6
Stockfish 18 safety gate
CompleteThe fixed-node root search starts with 37.Nb4 at about +6.81. Independent verification scores the unanimous 37.d8=Q+ model choice at about +8.39, so it passes the safety gate. The played 37.Qd5 falls to about +3.40.
- Engine move
- Nb4
- Depth
- 16
- Node budget
- 250,629
- Evaluation
- +6.81
Verification line
37.Nb4 Qxe5 38.Rxc6 Kb7 39.Rc1 a5 40.d8=Q Bxd8 41.Qxd8 axb4 42.Qc8+ Ka7 43.Rc7+ Qxc7 44.Qxc7+
Position lesson
Promotion with check is the clean move because Black must answer immediately. The new queen is exchanged, but White keeps the rook on the seventh rank and the exposed black king. A quiet queen move allows Black to attack the same pawn and reduce the pressure.
Interactive comparison
Choice three: underpromote with check
White to move after 37...Qxe5. Stockfish finds 38.d8=R+ and a forced mate. Every model takes the queen with 38.Qxe5, while the played 38.Rxc6 turns the position into a loss.
Model candidates
Choose a viewpoint
V6 Active: Qxe5
Measured GPU choice. The capture keeps a large advantage, but it misses the forced mate found by the engine.
Model top three: Qxe5 · d8=Q+ · d8=R+
Stockfish 18 safety gate
CompleteStockfish finds mate in fourteen with the rook underpromotion. The unanimous 38.Qxe5 model choice stays winning at about +5.50 but fails the mate-based gate. The played 38.Rxc6 changes the score to about -2.46.
- Engine move
- d8=R+
- Depth
- 29
- Node budget
- 250,713
- Mate
- M14
Verification line
38.d8=R+ Bxd8 39.Rc8+ Kb7 40.Nbc5+ bxc5 41.Nxc5+ Kb6 42.Nxa4+ Ka6 43.Rxc6+ Bb6 44.Nc5+ Ka5 45.Qxe5 Kb5 46.Rc8 Kc4 47.Na4+ Kd3 48.Nb2+ Kd2 49.Qc3+ Ke2 50.Re8+ Be3 51.Rxe3#
Position lesson
The rook promotion works because it gives check and removes the bishop on d8 after the forced reply. White then checks again with the rook on c8 and brings the knights into the attack. Taking the queen wins material, but it gives up the forced finish. Moving the rook from c7 loses even more.
Published finish
The forcing line
White had a winning attack after 34...Ba4. The game stayed favorable through 37.Qd5, but 38.Rxc6 let Black exchange queens, stop the passed pawn, and take over the ending.
Review method
Source, models, then verification
Start with the official public game
Every position comes from the complete Keetman-Van Arkel score linked by the Amsterdam Chess Open website.
Measure six saved evaluators
Each saved checkpoint ranks every legal move on the NVIDIA GPU with the same 30% hybrid positional weight.
Apply the Stockfish safety gate
Stockfish checks the engine best move, every distinct model choice, and the move played in the game at a fixed node budget.
What this review can and cannot show
Clear boundaries keep a useful experiment from becoming a bigger claim than the evidence supports.
- This review uses original teaching prose and does not reproduce third-party annotations.
- The official tournament page links the public broadcast. The reviewed game score identifies Group A, round 6, Apollohal in Amsterdam, and October 25, 2026.
- The six neural checkpoints are scalar evaluators, not move-policy, search, or explanation models.
- Their inputs encode piece placement only; move history, repetition, clocks, castling rights, and en-passant state are not model inputs.
- Profile labels describe intended training behavior, not proven human-like personalities or internal reasoning.
- Stockfish used 250,000 nodes per checked line, two threads, and 256 MB hash. This is a reproducible snapshot, not a claim of perfect play.
- Two-thread search can produce small score, depth, node-count, and principal-variation differences on another run.
- Forced-mate comparisons use the documented 100,000 cp value. Mate status and mate distance are shown separately.
- The prose describes visible board features and measured move consequences. It does not claim to reveal a model's internal reasoning.
Continue learning
Put the position on your own board
Start from the exact FEN and test whether you can find and convert the idea against a human-style bot.