From 60 to 63 Checking Moves: When AI Enters the World of Chess Composition

Nha Trang, October. Rain swept across the window. I was sorting through notebooks from my years in the VTC chess commentary booth when my son sent me a link: an AI system called Double Agent had just announced it solved a chess problem that had stood for 58 years. First read, I nodded. Second read, I put down my pen.

Something in the headline made me want to slow down. Over 32 years in the commentary booth, I learned one thing: a record is never old; it is only waiting for someone who knows how to listen. This time the listener was a machine, and the storyteller was a set of numbers lying still in a book.

The problem Double Agent solved is not a game. No clock, no opponent, no stands. It is a combinatorial problem — the kind the chess composition world has pursued for centuries: in a single position, how many legal checking moves can exist at most? And how long can a consecutive checking sequence run?

The question of checking moves has an answer. The prior record belonged to Norbert Ringeltaube with 60 checking moves. Double Agent produced a position reaching 63, and, more importantly, proved no position can exceed 63. A closed result: an example built, an upper bound blocked.

The remaining question is different. In 2026, Alexey Khanyan constructed a position with 54 consecutive checking moves. Double Agent offered 364 — but only as an upper bound, not the maximum. The problem stays open.

This is not a playground for players but for composers — a small, closed community where each record is written in pencil and checked by human eye. An empty stadium, yet the applause of 2026 still echoes in every old record. For nearly six decades they advanced step by step: one more checking move, then two, then three. Slow as a climber counting each stone.

The 58-year marker suggests a date. If the record was set in 2026, this report appeared around 2026 — an event both new and distant. It reminds me that some problems outlive the people who pose them.

The most interesting part is not the 63. It is how the machine got there.

The old way is brute force: build every possible position and count. The research team says plainly that this is infeasible — the configuration space is too large. Instead they followed a four-step path: prove geometric and structural properties of a "checking-everything" position; formalize those properties in Lean 4, a mathematical proof assistant; reduce the rest to a SAT problem; then let the machine check the whole proof.

The result comes not from computational power but from turning a chess intuition into a machine-checkable theorem. That is the methodological turning point.

One technical detail deserves a pause: the 63-move position must be legal — reachable from a real game, not an arbitrary "fairy" construct. The side to move must have at least one legal move, and every legal move must give check. To achieve that, the position must nearly compress piece density to its maximum, using pins and blocks to seal every non-checking move. This is the kind of thinking I have seen in composers: they do not add pieces, they tighten every empty square.

One point I want to make clear, because it is easily skimmed. The team suggests that adding pieces or firepower "may not improve the result." That is an opinion, not a proof, and the report presents it as such. In my trade, the distinction between "I believe" and "I can prove" is everything. A documentary can let a character say "I believe I once ran faster"; a theorem is not allowed to believe.

The difference between the two problems is the most beautiful part. The 63 problem is closed. The 364 problem is open. A machine proved an absolute limit and admitted a limit it has not touched. To a scientist, that admission is more trustworthy than any claim.

To see why this matters, recall that the composition world has no money, no audience, no prize. They work in silence, and each recognized record is a small honor. When a machine entered and broke Ringeltaube's 60-move record, it stole no one's glory. Every stadium I have walked through has a secret door, and the entrance is always a person. Here, that door opens onto a 63-move position, and the person behind it is a research team left unnamed.

I once sat listening to an elderly composer in Hue describe sleepless nights spent shifting a single position one square to the left. "Chess does not reward the fast," he said, "it rewards the patient." Reading Double Agent's result, I find him still right — only now the patient have an assistant that checks every step of reasoning.

In Vietnam, I know a few composers still working quietly. They have no Lean 4 assistant, no SAT problem. They have paper, pencil, and long nights. The gap between them and Double Agent is not a gap in talent — it is a gap in tools. And tools, unlike talent, can be learned.

And here I want to speak plainly.

The headline "AI solves a 58-year-old chess problem" is easily read as "AI has solved chess." Those are worlds apart. What was solved is a narrow combinatorial theorem — the maximum checking moves in one position — not chess as a game. Chess is not "solved." It is only understood one corner deeper.

A subtler risk: the entire verification was done by the research team itself. "Machine-checked" reduces the risk of error only if the Lean 4 formalization and the SAT reduction are themselves correct. A gap in the encoding of the problem is enough to collapse the upper bound. No independent audit is mentioned.

One more risk: the source is opaque. Most facts in the original report carry no specific attribution; only the headline figures are credited to Double Agent. To someone used to cross-checking across sources, this is a question mark to keep — not to deny the result, but to know where one stands.

I do not doubt the machine. I doubt how we tell its story. Over 32 years of commentary, I have seen people remember a chess move longer than a theorem — yet also forget that behind every result is a chain of choices that can be wrong.

What stays with me is not the 63 but the team's posture: they proved what could be proved and left open what could not. In an age when everyone wants to claim the summit, a machine that can say "I have not arrived" is the most trustworthy thing of all. Perhaps the better question is not "what has AI solved," but "when a machine knows where to stop, can people learn to stop there too?"