63 Legal Checking Moves and 364 Consecutive Checks: The Line Between Proof and Rumor
**Câu trả lời cốt lõi**: Hệ thống AI Double Agent đã chứng minh 63 là số nước chiếu hợp lệ tối đa trong một thế cờ, phá kỷ lục 60, đồng thời chứng minh 364 là giới hạn trên của chuỗi nước chiếu liên tiếp. Giá trị lớn nhất của bài toán thứ hai vẫn để ngỏ. **Dữ kiện chính**: - Double Agent công bố kết quả ngày 13 tháng 10 năm 2026, theo bản tin không nêu tên cơ quan công bố. - Kỷ lục cũ 60 nước chiếu thuộc về Norbert Ringeltaube; kết quả mới là 63, đã chứng minh là tối đa. - Kỷ lục cũ 54 nước chiếu liên tiếp thuộc về Alexey Khanyan, công bố năm 2013; kết quả mới là chặn trên 364. - Phương pháp gồm chứng minh cấu trúc, hình thức hóa trong Lean 4, quy về bài toán SAT và kiểm tra bằng máy. - Không có kiểm toán độc lập; toàn bộ chuỗi kiểm chứng do chính nhóm nghiên cứu thực hiện. **Nguồn**: Bản tin công bố ngày 13 tháng 10 năm 2026, cơ quan công bố không được nêu tên | Cross-checked: VuaBong.vn **Hỏi đáp liên quan**: - Hỏi: 63 có phải là số nước chiếu tối đa không? Đáp: Có, 63 đã được chứng minh là tối đa bằng cả chặn dưới và chặn trên. - Hỏi: 364 có phải là kỷ lục không? Đáp: Không, 364 chỉ là giới hạn trên; giá trị lớn nhất thực tế vẫn chưa được xác định. - Hỏi: AI đã giải được cờ vua chưa? Đáp: Chưa, kết quả chỉ là một định lý cực trị hẹp trong lĩnh vực sáng tác thế cờ.
On October 13, a short technical announcement appeared: an AI system named Double Agent claimed to have found a position with 63 legal checking moves — the maximum possible — and proved that no position can exceed that number. The previous record was 60, set by Norbert Ringeltaube. At the same time, the system produced a second number: 364 consecutive checking moves, but only as an upper bound, not the maximum. The old record in that category was 54, published by Alexey Khanyan in 2026.
Two numbers, two entirely different grades of evidence. And almost every report I read in the following two days merged them into a single story: “AI solves a chess problem.”
People called it a shock. I call it data that has not been read yet.
Context: this is a problem, not a game
One thing the headlines do not say needs saying immediately: there is no tournament here. No rated game, no player sitting at a board, no clock. The object of the problem is a very narrow branch of chess: problem composition and retrograde analysis. This is the work of people who design positions, not of people who compete.
The first problem asks: in a legal position, what is the maximum number of legal checking moves the side to move can have? The accompanying condition is strict: the side to move must have at least one legal move, and every one of those legal moves must give check. The second problem asks: how long can the longest sequence of consecutive checks be?
I have followed chess for more than thirty-seven years, but honestly, most of that time I looked at the board through the lens of match data. In 2026, at 45, I was one of only seven female data journalists accredited to the World Cup in Russia. In the Japan–Belgium match, I stayed behind after my colleagues had left the stand and wrote down: Japan generated 1.2 xG in the first 45 minutes, but after Fellaini came on in the 65th minute, Belgium won 14 of 18 aerial duels. A veteran editor dismissed my piece with the line: “Women looking at football data only know how to pick the numbers that suit them.” When it ran in a regional sports outlet, it drew more than 200,000 reads.
Since then I have set myself a rule: never write about any development without at least three clearly sourced numbers. When I moved to covering chess for the Chinese market, the rule did not change. Only the subject did.
And because of that rule, I have to separate this problem from any comparison with competitive chess.
The core: what was actually proven
The key point is this: 63 is a closed result; 364 is not.
For the first problem, the research team did two things at once. First, they constructed a position with exactly 63 legal checking moves — a lower bound. Second, they proved that no position reaches 64 — an upper bound. When the lower and upper bounds meet at the same number, the problem closes. 63 is the maximum. The old record of 60 by Norbert Ringeltaube is therefore surpassed by a margin of three moves.
For the second problem, the team completed only half. They proved that 364 is an upper bound: no sequence of consecutive checks is longer than 364. But they did not prove that a sequence of exactly 364 exists. The true maximum of this problem remains open. The best previously constructed record was 54, published by Alexey Khanyan in 2026 — and the jump from 54 to 364 is not a performance leap, it is a theoretical ceiling.
This is where most readers will be misled. A headline like “AI solves a 58-year-old chess problem” makes people picture something finished, complete. The reality is half finished, half hanging.
The genuine novelty of this work lies in the method, not the number. The team states plainly that they could not solve it by brute-force enumeration — the number of positions to consider is too large. Instead they followed a four-step pipeline: prove geometric and structural properties of checking configurations; formalize those proofs in Lean 4, a mathematical proof assistant; reduce the remainder to a SAT problem; then let the machine check the entire proof chain.
In other words, they did not search for the answer. They built a logical scaffold to prove the answer could not be otherwise.
The legality condition is a load-bearing requirement here, not decoration. The team asserts that their record position can arise from a legal game, not merely as a theoretical fairy construct. They also stress that the side to move must have at least one legal move, and that every legal move must give check. Three concepts — legal move, check, pinned piece — are all formally defined in the Laws of Chess. Without those definitions, the problem could not even be stated.
Structurally, the 63-position almost certainly maximizes piece density, using pins and blocks to suppress every non-checking move. The article itself mentions a tight arrangement, pins, line-blocking and square control. This is my inference, not the team's claim, but it fits the logic of the problem.
One notable timing detail: 58 years plus the 2026 record date gives 2026. That is an internal consistency point, and a reminder that this is a recent release, not a historical event.
The contrarian angle: machine-checked does not mean correct
There is a line I carry from football into chess: xG does not replace emotion, it explains why our hearts race. That principle applies here in a more uncomfortable way. Numbers do not replace belief in a result; they only show what that belief rests on.
And when I examined what this belief rests on, I found two gaps.
The first gap is the self-referential nature of the verification. The phrase “machine-checked” sounds very solid, but it reduces the risk of error only on one condition: that the formalization itself is correct. If the SAT reduction has an encoding gap — an under-specified constraint — the upper bound can collapse. The article mentions no independent audit. The entire verification chain sits with the research team itself.
The second gap is source quality. The publishing outlet is not named. Most of the facts in the original report have no source — only the headline numbers are attributed, to Double Agent. For an extremal result, the absence of independent verification is the single biggest risk, and here it has not been addressed.
One claim also needs to be separated from the proof. The team says that adding pieces or attacking power “may not improve the result.” That is an opinion, not a theorem. The article presents it as such — but a skimming reader will not distinguish proof from conjecture.
Then there is the gap between headline and content. The headline suggests chess has been “solved.” In reality a narrow extremal theorem has been proven, and another question remains open. This is a familiar distortion: the top of the article travels far beyond the body. I saw exactly this mechanism in the Evergrande piece in 2026. When I predicted that club would collapse on fitness — average sprint speed down 38.2 percent, players over 30 recovering 27.3 percent more slowly after a six-month pandemic break — the editorial board laughed. On December 12, 2026, Evergrande lost 1–2 in the final, running 6.4 km less than their opponents. The data was right, but only if you read correctly what it was talking about.
The takeaway
Numbers are asceticism: you must give up what is easy before you can see what is true.
For the 63-move problem, the data is sufficient to conclude. It is a closed extremal theorem, with lower and upper bounds meeting, proven through a formalized pipeline. Its industry value lies elsewhere: it normalizes machine-checkable proof workflows in one small corner of the chess composition world. Over time, that may be the most durable effect — not the number 63, but the verification standard it imposes on future extremal claims.
For the 364 problem, the data is not sufficient. It is an upper bound, not a maximum. Anyone citing it as “AI has finished solving it” is misreading the original text.
I am staking a falsifiable claim: if in the next six months no group independently reruns the Lean 4 chain and the SAT encoding, or if a gap in the SAT reduction is shown, then the 364 upper bound — and possibly 63 as well — will have to be revisited. The signal to watch in the next cycle is not headlines saying “AI solves chess,” but whether the team publicly releases its proof files and source code. An extremal theorem is only credible when others can rerun it. Until then, I keep my conclusion: 63 is a proven fact, 364 is a promise not yet fulfilled.
