Double Agent finds chess position where all 63 moves check, beating 1968 record Double AI's Double Agent found a chess position, White to move, in which all 63 legal moves give check, surpassing the record of 60 that had stood since 1968, and proved that 64 is impossible. Double Agent also constructed a real 190-ply game from the starting position that reaches the position, using pruning filters, pawn-history reconstruction and a backtracking search guided by a 17-move plan. The result extends the company's work on extremal combinatorics problems. Blog · Nov 2026 Can LLMs redesign themselves as they train? https://www.doubleai.com/research/can-llms-redesign-themselves-as-they-train Architecture is the last hand-designed piece. We make it part of training. Double Agent found a chess position in which all 63 legal moves give check, beating a record that had stood since 1968. Date Oct 19, 2026 Type Blog Research area Publications Reading time 5 min How many legal moves can a chess position have if every single one of them gives check? Since 1968 the record was 60. 1 Double Agent found a position, White to move, with 63, and proved that 64 is impossible. It also found a real game from the starting position that reaches it. Given the enormous state space of chess https://en.wikipedia.org/wiki/Shannon number , this takes a combination of high-level chess reasoning and exhaustive computer search https://www.doubleai.com/about : Extremal combinatorics https://en.wikipedia.org/wiki/Extremal combinatorics asks how far you can push something before it breaks. How many friendships can a group of people have before some three of them are all friends with each other? How many numbers can you pick from 1 to 1,000 before three of them are evenly spaced? The objects that sit right at the edge are called extremal constructions. We have pointed Double Agent at problems like these in graph theory https://www.doubleai.com/research/advancing-a-decades-old-problem-in-extremal-graph-theory before. Chess is a beast of its own, and composers have been setting records in it for over a century. This time we chose one of them: a position in which every legal move gives check, with as many legal moves as possible. 2 White to move. Each of White's 63 legal moves gives check, and 23 of them are mate. Position on Lichess · the 190-ply game Finding the position and finding a game that reaches it are intertwined. Searching over all positions would be pointless, because most of them are blatantly illegal: So Double Agent prunes as it searches. Further filters reconstruct the history of the pawns. Pawns are predictable: they only move forward, and they change file only when they capture. For each candidate position, Double Agent checks that at least one history makes sense: some way of assigning pawns to starting files, with captures and promotions that add up. Three examples of positions with no such history: With these filters in place, only one placement of the kings can still reach 63: Black on c6, White on d3. With a likely position in hand, Double Agent has to find an actual game that ends there. Off-the-shelf proof-game solvers such as Texel https://github.com/peterosterlund2/texel get stuck on crowded positions like this one: they estimate how many moves remain by assuming pieces never get in each other's way. So Double Agent combines high-level planning with backtracking search. It first decides which captures and promotions the game must contain, and roughly when. The search then only has to fill in the moves between those milestones. The position makes strict demands: White's five captures each do more than one job: Double Agent wrote this plan out as the first 17 moves: 1.a3 b5 2.c4 b4 3.axb4 d5 4.cxd5 h5 5.g4 e5 6.gxh5 f5 7.Nf3 c5 8.Nxe5 c4 9.Nf3 g5 10.Nd4 c3 11.Nxf5 g4 12.d3 c2 13.Bd2 g3 14.Bc3 g2 15.e4 g1=Q 16.Nd2 Qg4 17.Be2 c1=Q Backtracking search, guided by the plan, then fills in the remaining 160 or so plies, walking five new queens and Black's pieces into place. Full game on Lichess https://lichess.org/analysis/pgn/1.a3 b5 2.c4 b4 3.axb4 d5 4.cxd5 h5 5.g4 e5 6.gxh5 f5 7.Nf3 c5 8.Nxe5 c4 9.Nf3 g5 10.Nd4 c3 11.Nxf5 g4 12.d3 c2 13.Bd2 g3 14.Bc3 g2 15.e4 g1%3DQ 16.Nd2 Qg4 17.Be2 c1%3DQ 18.b5 Kd7 19.h4 a5 20.d6 Qgg5 21.b6 a4 22.h6 Qg4 23.b4 Qf6 24.b7 Be7 25.e5 Ra5 26.e6%2B Kc6 27.d7 Ra7 28.d4 Qfg5 29.h5 Qg3 30.h7 Kb6 31.h6 Q3h4 32.d5 Qh2 33.d6 Qa3 34.Rc1 Nc6 35.Ne4 Bf8 36.e7 Qb2 37.Bd2 Na5 38.Rc2 Qg4 39.Bf3 Qb1 40.Ke2 Nc6 41.Nc3 Qb3 42.Rg1 Qd4 43.Bh1 Nf6 44.Rg2 Ne8 45.Qg1 Na5 46.b8%3DQ%2B Bb7 47.Ng3 Qg4%2B 48.Kd3 Qb1 49.Bf4 Nb3 50.Bg5 Rg8 51.h8%3DQ Ra6 52.h7 Bh6 53.Qe5 Qgd1%2B 54.Bd2 Rg4 55.f3%2B Be3 56.f4 Qh5 57.Qf2 Rh4 58.Qg1 Rh3 59.Qf2 Ra7 60.Qg1 Ka6 61.Qf2 Bd5 62.Qg1 Rb7 63.Qf2 Kb6 64.Qg1 Be6 65.Qa5%2B Kc6 66.h8%3DQ Rb6 67.Qhe5 Bg8 68.Nf5 Qh7 69.Qf2 Rb7 70.Qg1 Rc7 71.Qf2 Rc8 72.Qg1 Ng7 73.Qf2 Rf8 74.Qg1 Qdh5 75.Bc1 Ne8 76.Qba7 Rh4 77.Qb8 Bd2 78.Qga7 Be3 79.Bd2 Rh3 80.Bc1 Rf6 81.Bd2 Nd4 82.Bc1 Nf3 83.Bd2 Bf2 84.Bc1 Be1 85.Be3 Qg4 86.d8%3DQ Bd2 87.d7 Ng7 88.Qe8 Bc1 89.Qf7 Ba3 90.d8%3DQ Ne8 91.Qed4 Nd2 92.Qe6%2B Nd6 93.Qee5 Ne8 94.Rf2%2B Qf3 95.Rg2 Qfd1 We prove the impossibility using SAT certificates. The idea is to turn the question into a giant logic puzzle, whose variables are yes/no questions: The rules of chess become conditions linking them: On top of the rules goes the goal, "at least 64 legal moves, all of them checks", which needs many more conditions, since the solver has to count. A SAT solver https://en.wikipedia.org/wiki/Boolean satisfiability problem then searches for an assignment satisfying all of them. We split the problem into 7,224 cases, by the squares of the two kings and whether White is in check. Every case is unsatisfiable. For each case the solver produced an LRAT proof, which was verified by independent proof checkers. Together the proofs come to about 5 TB and took 126 core-hours to produce. 1 Per the one-mover records register Einzügerrekorde http://www.skcaissa.de/einzueger.htm , maintained by Max Witte. 2 The position must be reachable: an actual legal game from the starting position has to end there. A promotion counts as four moves, one per piece, as is standard. Keep reading The first Code Execution Model. An agent of the machine.