White 321 vs Black 672
White starts from setup 321 (BNRBQKRN), Black from setup 672 (BBQRKNNR). In the encyclopedia since 2026-10-15: neither side starts better.
|cp| ≤ 5 at 400M · 800M · 1.6B · 4B nodes, two independent engines agree
How to play with different setups
White sets up the pieces by position No. 321, Black by position No. 672, mirrored on their own side; the pawns stand as usual.
Moves as usual: White starts; everything else follows the Chess960 rules.
The two setups
White: setup #321 B N R B Q K R N
King on the f-file, rooks on c, g, queen on the e-file; one bishop starts in a corner; the knights start away from the centre. Castling follows the Chess960 rule: king and rook end on the classical squares. In ordinary symmetric play setup 321 gives White +0.27 (Morozov Balance Index, 400M nodes).
Black: setup #672 B B Q R K N N R
King on the e-file, rooks on d, h, queen on the c-file; one bishop starts in a corner; the knights start on central files. Castling follows the Chess960 rule: king and rook end on the classical squares. In ordinary symmetric play setup 672 gives White +0.26 (Morozov Balance Index, 400M nodes).
Why the pair is balanced
At the deepest rungs (400M, 800M, 1.6B nodes) the evaluation is exactly 0.00 each time, reaching search depth 55.
Admission rule: |cp| ≤ 5 at four budgets up to 4 billion nodes, no one-sided drift along the ladder, and confirmation by two independent engines — Berserk and Lc0 — each below its own median on symmetric positions. The same corridor is required at every deeper budget (8 G, 16 G) at which the pair has been evaluated.
The shallow rungs agreed early: from 100 thousand to 128 million nodes the evaluation never left ±13 cp.
For comparison, in symmetric play setup 321 gives White +0.27 and setup 672 gives White +0.26: with both sides on the same rank the first move is worth something. In this pair Black's own setup absorbs that edge completely.
White is not forced into a single path: the best first move is b4, and 3 first moves are within 10 cp of it.
Along the engine's main line (30 moves) the largest swing was 46 cp (around move 8), ending at 0 cp — the position stays playable for both sides.
Engine self-play agrees: 10 games, White won 0, drawn 10, Black won 0 (White scored 50.0%).
Verification ladder
Nine independent searches of the starting position, each with a larger node budget. Evaluation in centipawns from White's point of view.
| Stage | Nodes | Depth | Eval, cp | Main line |
|---|---|---|---|---|
| S0 | 100 000 | 15 | -6 | 1. b4 Nf6 2. O-O b6 3. c4 Qb7 4. f3 c5 |
| S1 | 500 000 | 21 | -1 | 1. O-O c5 2. c4 Nf6 3. e3 b6 4. f4 Ng6 |
| S2 | 2 000 000 | 24 | -9 | 1. b4 Nf6 2. O-O c5 3. bxc5 Qxc5 4. e3 b5 |
| S3 | 8 000 000 | 29 | +13 | 1. b4 Nf6 2. O-O c5 3. bxc5 Qxc5 4. e3 b5 |
| S4 | 32 000 000 | 39 | 0 | 1. O-O Nf6 2. b4 d5 3. e3 b6 4. c4 dxc4 |
| S5 | 128 000 000 | 42 | +6 | 1. b4 Nf6 2. O-O d5 3. e3 b6 4. c4 dxc4 |
| S6 | 400 000 000 | 46 | 0 | 1. b4 Nf6 2. O-O d5 3. e3 b6 4. c4 dxc4 |
| S7 | 800 000 000 | 53 | 0 | 1. f4 Nf6 2. c4 c5 3. e3 Ng6 4. Nc3 b6 |
| S8 | 1 600 000 000 | 55 | 0 | 1. f4 Nf6 2. c4 c5 3. e3 Ng6 4. Nc3 b6 |
Deep verification
The search ladder reaches 1.6 billion nodes; admission needs a 4-billion rung and two further engines on top of that. Beyond even that, every pair was re-checked at budgets up to ten times larger and in thousands of self-play games. This section shows what each check returned for this pair.
| Nodes | Eval, cp |
|---|---|
| 4 000 000 000 | 0 |
| 8 000 000 000 | not checked |
| 16 000 000 000 | not checked |
This pair stays within the ±5 cp corridor up to 4 000 000 000 nodes.
Independent engines, run on the same position:
- Berserk 14: -1
- Lc0 0.32.1: -17
Independent confirmations: 3 of 3. An engine counts as confirming when its reading of the pair is closer to equality than its own reading of a typical symmetric Chess960 position (+10 for Berserk, +18 for Lc0). Where only one engine confirms, the pair is balanced according to that engine alone, and this page says so.
Own weight of each setup, measured against the classical array: White +86, Black +156, difference -70. A balanced pair needs White's setup to be weaker by about the price of the first move.
Extended self-play: 10 games, 0 White wins, 10 draws, 0 Black wins, White scored 50.0%.
Best first moves
The 3 best first moves for White, searched together at 400 000 000 nodes.
| # | Eval, cp | Depth | Main line |
|---|---|---|---|
| 1 | -4 | 45 | 1. b4 Nf6 2. O-O d5 3. e3 b6 4. c4 dxc4 5. Rxc4 Qa6 |
| 2 | -8 | 45 | 1. O-O Nf6 2. b4 d5 3. e3 b6 4. c4 dxc4 5. Rxc4 Qa6 |
| 3 | -8 | 45 | 1. e3 Nf6 2. b4 b6 3. O-O d5 4. c4 dxc4 5. Rxc4 Qa6 |
Main-line play-out
30 moves, 5 000 000 nodes per half-move. Largest |evaluation| 46 cp, mean 8.4 cp, final 0 cp.
| # | White | cp | Black | cp |
|---|---|---|---|---|
| 1 | b4 | +12 | Nf6 | +19 |
| 2 | O-O | +22 | d5 | +1 |
| 3 | e3 | +11 | b6 | 0 |
| 4 | c4 | 0 | dxc4 | 0 |
| 5 | Rxc4 | 0 | Ne6 | 0 |
| 6 | d4 | -4 | h5 | 0 |
| 7 | f4 | -33 | h4 | -32 |
| 8 | Nd2 | -40 | c5 | -46 |
| 9 | bxc5 | -41 | bxc5 | -22 |
| 10 | Bf3 | -16 | Bxf3 | -21 |
| 11 | Nxf3 | -16 | Qa6 | -18 |
| 12 | Rc2 | -16 | cxd4 | -15 |
| 13 | Nxd4 | -14 | Nxd4 | -14 |
| 14 | Bxd4 | -17 | g5 | -19 |
| 15 | Qe2 | -10 | Qxe2 | -13 |
| 16 | Rxe2 | -3 | gxf4 | -6 |
| 17 | Bxf6 | -6 | exf6 | -5 |
| 18 | exf4+ | 0 | Kf8 | 0 |
| 19 | g3 | 0 | Bc7 | 0 |
| 20 | Kg2 | -2 | Bb6 | -4 |
| 21 | Nf2 | -4 | Kg7 | -1 |
| 22 | Rb1 | 0 | Rd4 | 0 |
| 23 | Rd1 | 0 | Ra4 | 0 |
| 24 | Rd3 | 0 | f5 | 0 |
| 25 | Rd5 | 0 | Kg6 | 0 |
| 26 | Rd6+ | 0 | Kg7 | 0 |
| 27 | Rd5 | 0 | Kg6 | 0 |
| 28 | Rd6+ | 0 | f6 | 0 |
| 29 | Nd3 | 0 | Kf7 | 0 |
| 30 | Rd7+ | 0 | Kg6 | 0 |
Engine self-play
10 games at 500 000 nodes per move: White won 0, drawn 10, Black won 0 — White scored 50.0%. Average length 37.3 moves.
Sample game 1 (1/2-1/2)
1. e3 Nf6 2. O-O b5 3. b4 Ne6 4. f4 c5 5. bxc5 Qxc5 6. d4 Qc7 7. Ng3 h5 8. h4 g6 9. a4 a6 10. axb5 axb5 11. Nc3 O-O 12. Nxb5 Qc6 13. Bf3 Qxb5 14. Bxa8 Bxf4 15. exf4 Rxa8 16. f5 Ng7 17. Qxe7 Ng4 18. Rb1 Qa5 19. Qxd7 Ne3 20. Rfe1 Ngxf5 21. Nxf5 Nxf5 22. Qb5 Qxb5 23. Rxb5 Rfc8 24. Rc5 Rxc5 25. dxc5 Kf8 26. Bf6 Ra6 27. Bc3 Ra4 28. Bf6 Ra6 29. Bc3 Ra4 30. Bf6
Sample game 2 (1/2-1/2)
1. e3 Nf6 2. c4 b6 3. f4 c5 4. b4 cxb4 5. O-O Ng6 6. d4 O-O 7. f5 Qc7 8. g3 Nh8 9. Qxb4 d5 10. Nd2 e6 11. fxe6 fxe6 12. Bb2 dxc4 13. Qxc4 Qxc4 14. Nxc4 Nf7 15. Nf2 Ng5 16. Ba3 Rfe8 17. Nd2 a6 18. h4 Nf7 19. Nc4 b5 20. Nb6 Bb7 21. Nd3 Nd6 22. Nc5 Bd5 23. Nxa6 Nc4 24. Nxc4 Bxc4 25. Nxb8 Bxf1 26. Na6 Ra8 27. Nc7 Bc4 28. Nxe8 Nxe8 29. Rc3 Nf6 30. Bf3 Nd5 31. Bxd5 exd5 32. e4 dxe4 33. Kf2 Kf7 34. Ke3 Ra4 35. Kxe4 Ke6 36. Bf8 Rxa2 37. Re3 Kd7 38. Bc5 Kc6 39. Ra3 Bd5+ 40. Ke5 Re2+
Play it in the CH960 app
Open this page on a phone with the CH960 app installed, or find pair 321 / 672 in the app's encyclopedia, then choose how to play.
- Local game — two players on one device, with a clock.
- With a mentor — a clearly labelled training bot with hints, at your level.
- Challenge a friend — an online game by link; both players get the same pair, colours swap on rematch.
Play this pair against the computer, right here Play a friend
FEN (Shredder, castling rights by rook files): bbqrknnr/pppppppp/8/8/8/8/PPPPPPPP/BNRBQKRN w CGdh - 0 1
Mirror twin: White 347 vs Black 479 — a different position (castling squares are fixed), not evaluated by the search.