Tanoi vs Ker: Open Files and Active Rooks
Three decisions show how a central capture and a pawn break make room for active rooks.
Reproducible analysis run complete
6 saved models on NVIDIA GeForce RTX 5080 · Stockfish 18 · 250,000 nodes per checked line.
Why this game?
Look beyond the first capture
A strong move can change what the pieces are able to do next. In this Wellington Open game, Anthony Ker challenges the centre, opens a file and brings his rooks into White's position. Toleafoa Edward Tanoi's replies make each decision worth exploring. Start with Black's choices on moves 19, 28 and 36. At each stop, look for the opponent's reply before trusting a tempting capture or a check.
- White
- Toleafoa Edward Tanoi
- Black
- Anthony F Ker
- Event
- Wellington Easter Open, round 1, Wellington, New Zealand
- Result
- 0-1
Interactive comparison
Choice one: look behind the central pawn
Black to move after 19.Qxa3. Anthony Ker played 19...Nxd5. Before taking back with the e4-pawn, notice what that pawn is shielding on the e-file.
1r2r1k1/1pqb1pp1/p2p1nnp/2pP4/P3P3/Q4P2/NP2B1PP/R1B2RK1 b - - 0 19Model candidates
Choose a viewpoint
V6 Active: Bxa4
Measured GPU choice, rejected by Stockfish. Explore it to see the tactical problem, not as a recommended move.
Model top three: Bxa4 · Rxe4 · Nxe4
Stockfish 18 safety gate
CompleteFrom Black's view, Stockfish selects the played ...Nxd5 at +1.17. None of the three distinct model choices passes the 100-centipawn gate. Their agreement is not a reason to play an unsound capture.
- Engine move
- Nxd5
- Depth
- 15
- Node budget
- 250,252
- Evaluation
- +1.17
Verification line
19...Nxd5 20. Rd1
Position lesson
The rook on e8 faces White's bishop on e2, with the e4-pawn between them. If White answers ...Nxd5 with exd5, the rook can capture that bishop. A defended pawn is not always safe when its defender also protects another piece. The tempting model captures on a4 and e4 do not survive the engine check.
Interactive comparison
Choice two: open a route for the rooks
Black to move after 28.Qb2. The game continued 28...c4 29.bxc4 bxc4. Compare this pawn break with the forcing-looking ...Nxf3+.
2r1r1k1/3nq1p1/p2pp2p/1pp1n3/P3P3/1P3P2/NQ3BPP/2RR2K1 b - - 1 28Model candidates
Choose a viewpoint
V6 Active: Nxf3+
Measured GPU choice, rejected by Stockfish. Explore it to see the tactical problem, not as a recommended move.
Model top three: Nxf3+ · bxa4 · c4
Stockfish 18 safety gate
CompleteStockfish selects the played ...c4 at +1.95 for Black; its separate forced-line check returns +1.84. V17's ...bxa4 also passes at +1.84, only 11 centipawns below the unrestricted reference. The five-model ...Nxf3+ choice fails at -0.98.
- Engine move
- c4
- Depth
- 15
- Node budget
- 250,061
- Evaluation
- +1.95
Verification line
28...c4
Position lesson
The exchanges clear the b-file and leave a black pawn on c4 controlling d3. After White's queen moves to d4, ...Nd3 attacks the rook on c1 and bishop on f2. Follow the pawn exchanges before judging the break: their value is the space and routes they create, not just material collected.
Interactive comparison
Choice three: check before collecting material
Black to move after 36.Rxd6. All six evaluators suggest ...Qxd6, but White's queen on a3 can recapture along a3-b4-c5-d6. Ker instead played 36...Rb1+.
1r4k1/2q3p1/p2Rp2p/4n3/P3P3/Q4P1P/N1r2BP1/6K1 b - - 0 36Model candidates
Choose a viewpoint
V6 Active: Qxd6
Measured GPU choice, rejected by Stockfish. Explore it to see the tactical problem, not as a recommended move.
Model top three: Qxd6 · Rxf2 · Rxa2
Stockfish 18 safety gate
CompleteThe unrestricted search chooses ...Nc4 at +7.00 for Black. A separate check of the played ...Rb1+ scores +7.65 and passes. This finite search does not establish a reliable ordering between those winning moves. The unanimous model choice ...Qxd6 scores -1.79 and is rejected.
- Engine move
- Nc4
- Depth
- 16
- Node budget
- 250,511
- Evaluation
- +7.00
Verification line
36...Nc4 37. Qd3 Nxd6 38. Bg3 Qc5+ 39. Kh2 Rbb2 40. Qf1 Rxa2 41. Bh4 Rxg2+ 42. Qxg2 Qe5+ 43. Bg3 Rxg2+ 44. Kxg2 Qb2+ 45. Bf2
Position lesson
Do not stop after seeing an undefended rook. Count the next capture too: ...Qxd6 Qxd6 gives up the queen for that rook. Black can keep the queen and use the open b-file instead. After ...Rb1+ 37.Nc1 Rbxc1+ 38.Kh2 Rxf2, the rooks work together and Black remains decisively ahead. This is an analysed continuation, not extra moves from the recorded game.
Published finish
The forcing line
This is the recorded continuation from Black's move 28 to the end of the public score. The broadcast gives 0-1 after 36...Rb1+; it does not record a checkmate or a reason the game ended. The earlier 19...Nxd5 decision is shown separately.
Review method
Source, models, then verification
Replay the public game
Use the broadcast linked from Wellington's official event page. Replay all 72 plies legally 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 are unchanged.
Check each distinct choice
Stockfish 18 checks its best line, each distinct model top move 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.
- The public broadcast identifies this game as round one on 3 April 2026. Wellington's official event page confirms the event dates; the broadcast supplies the players, moves and result.
- Only the factual game score is reused. These explanations are original; broadcast annotations, clock comments and evaluations are not copied.
- All displayed evaluations use the side-to-move perspective. Every reviewed position is Black to move, so positive scores favor Black.
- The 100-centipawn gate is a broad safety threshold, not a promise of best play. Only one distinct model move passes across these three positions.
- At move 36, the independent forced ...Rb1+ search scores higher than the unrestricted ...Nc4 reference. The generator clamps negative loss to zero; do not infer a definitive ranking from this limited search.
- 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 behavior 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 was required to use two search threads. Parallel search can produce small score, depth, node-count, or PV differences on a fresh rerun even when the configured node limit is unchanged.
- Centipawn loss uses a 100,000 cp comparison value for forced-mate scores; mate status and mate distance are also reported separately.
- The analysis does not measure Elo, playing strength, or clean unseen test accuracy.
- Any reader-facing move explanation must describe observable chess features; it must not be presented as the 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.