How the Suspicious Move Score Is Calculated

The four signals

For a given move, the score is built from four independent measurements. Each captures a different facet of “looks like a solver, not a person,” and each is reported alongside the final score so a result can be inspected rather than taken on faith.

Solver Match

Solver Match asks how closely the observed move tracks what a strong game solver would choose from the same position. The position is reconstructed as it stood before the move, a solver evaluates the legal plays, and the observed move is located within that ranking — how near the top it sits, and how little scoring value and board equity it gives up relative to the best available option. A move that is repeatedly the solver's first or near-first choice scores high; an ordinary, good-enough move does not.

Move Difficulty

The same strong move means very different things on an easy board and a hard one. Move Difficulty estimates how demanding the position was: whether the best play was obvious or buried among many candidates, how crowded the board had become, and how few near-optimal options existed. Finding the one non-obvious best move on a genuinely difficult position is far more distinctive than making the obvious play everyone would find.

Word Unusualness

Word Unusualness measures how rarely the played word appears in real human play, taking the player's skill level into account. Every word the score considers is legal — this is not about validity — but solvers and word-finders have the whole dictionary memorized and readily surface obscure, seldom-used entries that a person would rarely reach for. An unusual word is more surprising coming from a human, so it raises this signal.

Rack Robustness

A move can look brilliant under one lucky assumption about the opponent's tiles and unremarkable under another. Rack Robustness measures how consistently the move looks strong across the range of racks the opponent could plausibly have been holding. A play that is near-optimal for many plausible racks — not just one convenient guess — is a sturdier signal than one that depends on a single assumed hand.

Hidden-rack sampling

In a real game you almost never know your opponent's exact tiles, and their rack changes what “the best move” even is. Rather than assume one hand, the method samples plausible racks that are consistent with everything visible — the tile bag minus the tiles already seen on the board and in the move. It evaluates the move under each sampled rack and weights the results by how likely each rack is.

This is why several of the signals above are described “across the racks the opponent could have held.” Sampling keeps the score honest about uncertainty: it reflects a distribution of possible situations instead of a single convenient assumption, and it is what lets Rack Robustness exist as a measurement at all.

How the signals combine

The four signals are not simply averaged. Solver Match acts as a gate: a move that is not solver-like cannot earn a high score no matter how unusual its word or how hard the board was. Suspicion is fundamentally about optimal play, so nothing else can substitute for it.

Once a move is solver-like, the other three signals raise the score. A move that closely matches the solver, was hard to find, uses an unusual word, and holds up across many plausible racks is precisely the combination that unaided human play rarely produces — so it lands high on the scale. A solver-like but obvious move on an easy board with a common word stays modest. The signals are then mapped to a single 0–100 number with a plain-language band, calibrated against how human players score on the same measurements.

One note on calibration: the human reference is drawn from real games, and where a game does not yet have its own human dataset, a shared reference is used and clearly labeled in the result. Calibration is strongest within the range the reference data actually supports, and scores beyond that range are shown with an honest note rather than implied precision.

What the score does not mean

The Suspicious Move Score is a measurement of how unusual play is — not a detector of cheating. It is worth being precise about its limits:

• It is not proof. It cannot see a second device, a word-finder, or anyone's intent. It only compares play against solver and human references.
• One move is weak evidence. Luck and skill both produce excellent individual moves constantly. A single high score is a reason to look closer, never a conclusion.
• Skill is not cheating. Strong players naturally make solver-like moves more often and know more unusual words. The question is whether a pattern exceeds what skill and luck reasonably explain.
• A pattern is what matters. Repeated high scores on non-obvious, maximal moves across many turns are far harder to explain than any one result, which is why analyzing several moves beats fixating on one.

Used this way — as one analytical input, read across a body of play rather than a single screenshot — the score is a useful guide to where to look more closely. Treated as a verdict, it will mislead you.