Yang vs Gao: Defend Before the Attack Lands
Three decisions show why a tempting capture can fail and why king safety must come before material.
Reproducible analysis run complete
6 saved models on NVIDIA GeForce RTX 5080 · Stockfish 18 · 250,000 nodes per checked line.
Why this game?
Check the threat before taking material
Ray Yang and Arthur Gao reach a sharp Sicilian position where both kings and several loose pieces demand careful calculation. The first question is whether White can capture an active knight. The next two ask Black to choose between material and defence. At each stop, trace every attacking line before trusting the most tempting capture.
- White
- Ray Yang
- Black
- Arthur Gao
- Event
- Victorian Championship Premier, round 1, Melbourne, Australia
- Result
- 1-0
Interactive comparison
Choice one: do not grab the attacking knight
White to move after 22...Ng4. The knight looks loose, but first trace the long diagonal from Black's bishop on c8 to g4.
1rb2r1k/7p/1q1p1p2/p2B2b1/R4Nn1/1P6/2P3PP/1N1Q1R1K w - - 1 23Model candidates
Choose a viewpoint
V6 Active: Qxg4
Measured GPU choice, rejected by Stockfish. The bishop on c8 guards g4, so this queen capture loses material immediately.
Model top three: Qxg4 · Rxa5 · Nd3
Stockfish 18 safety gate
CompleteStockfish gives Black a small edge after 23.Qf3. The played 23.Qd2 is only 29 centipawns behind in its separate check, while 23.Qxg4 falls 371 centipawns behind the best line.
- Engine move
- Qf3
- Depth
- 15
- Node budget
- 250,010
- Evaluation
- -0.76
Verification line
23.Qf3 Bd7
Position lesson
All six evaluators choose 23.Qxg4, but 23...Bxg4 simply takes the queen. Stockfish slightly prefers 23.Qf3. Ray Yang played 23.Qd2, which also keeps the queen safe and stays close to the engine choice.
Interactive comparison
Choice two: take before retreating
Black to move after 25.h4. The bishop on g5 can capture the knight on f4, while the bishop on d7 can take the rook on a4.
1r3r1k/3b3p/1q1p1p2/p2B2b1/R4N1P/1P2n3/2PQ2P1/1N2R2K b - - 0 25Model candidates
Choose a viewpoint
V6 Active: Bxa4
Measured GPU choice. Stockfish places this capture only 19 centipawns behind its best move, so it passes the stated safety gate.
Model top three: Bxa4 · Nxd5 · Bxf4
Stockfish 18 safety gate
CompleteStockfish slightly prefers 25...Bxf4. The six-model choice 25...Bxa4 also passes at just 19 centipawns behind. The played retreat 25...Bh6 loses 152 centipawns and turns the balance toward White.
- Engine move
- Bxf4
- Depth
- 13
- Node budget
- 250,366
- Evaluation
- +0.12
Verification line
25...Bxf4 26.Rxf4
Position lesson
Both captures address a concrete target. Stockfish prefers 25...Bxf4, and every model chooses the safe 25...Bxa4. The retreat 25...Bh6 leaves the knight on e3 exposed to 26.Rxe3 and gives White the initiative.
Interactive comparison
Choice three: defend the king first
Black to move after 27.Re7. White's rook has reached the seventh rank and the queen can soon join the attack from f4 and h6.
1r3r1k/4R2p/1q1p1p1b/p2B4/b4N1P/1P6/2PQ2P1/1N5K b - - 1 27Model candidates
Choose a viewpoint
V6 Active: Bxf4
Measured GPU choice and the move played in the game. Stockfish rejects it because king safety matters more than collecting the knight.
Model top three: Bxf4 · Qxb3 · Bxb3
Stockfish 18 safety gate
CompleteWhite is already better, but 27...Be8 offers the strongest defence in this fixed-node check. The played 27...Bxf4 falls another 277 centipawns behind and allows the queen to join the attack.
- Engine move
- Be8
- Depth
- 14
- Node budget
- 250,226
- Evaluation
- -2.09
Verification line
27...Be8 28.Ng6+
Position lesson
The immediate task is king safety. Stockfish prefers 27...Be8, moving the a4-bishop back to challenge the rook and strengthen the dark squares. The played 27...Bxf4 wins a knight but lets White's queen enter with force.
Published finish
The forcing line
This is the recorded continuation from White's move 23 to the end of the public score. The broadcast gives 1-0 after 29.Qh6; it does not record checkmate or state why the game ended.
Review method
Source, models, then verification
Replay the public game
Use the championship broadcast supplied by Melbourne Chess Club. Replay all 57 legal plies and choose three teaching positions before measuring the models.
Measure six saved evaluators
Rank every legal move on the RTX 5080 with the fixed 30% hybrid positional weight. The existing checkpoints remain unchanged.
Check each distinct choice
Stockfish 18 checks its best line, every distinct model choice, and the played move independently with 250,000 nodes per line, two threads, and 256 MB hash.
What this review can and cannot show
Clear boundaries keep a useful experiment from becoming a bigger claim than the evidence supports.
- Melbourne Chess Club's official event page lists round one on 22 August 2026, while the public broadcast PGN header gives 26 August 2026. This review does not guess which date is correct.
- The broadcast identifies the players, round, moves, and result. Only the factual game score is reused; the teaching explanations are original.
- The public score ends with 1-0 after 29.Qh6 and does not state whether Black resigned or another finish was omitted.
- All displayed evaluations use the side-to-move perspective. Positive values favour the player whose turn it is in the reviewed position.
- The six neural checkpoints are scalar evaluators, not move-policy models or explanation models.
- Their 768 inputs encode piece placement only; side to move is applied after inference, while castling rights, en-passant state, repetition, move history, and clocks are not model inputs.
- Profile labels describe intended training or runtime behaviour and do not prove human-like personality or internal understanding.
- Stockfish results use a fixed node budget and configuration and are a verification snapshot, not a claim of perfect play.
- Stockfish used two search threads. Parallel search can produce small score, depth, node-count, or principal-variation differences on a fresh rerun.
- The analysis does not measure Elo, playing strength, or clean unseen test accuracy.
- The prose explains observable chess features; it does not claim to reveal internal model 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.