2048

In a row of four 2s, one move produces two 4s — not a single 8. A tile created by a merge is spent for the rest of that move, and implementations that miss this rule are dramatically easier than the real game.

It is the single most commonly broken rule in browser clones, and it is worth checking before you trust a score from anywhere. Line up four identical tiles, make one move, and count what comes back. Two tiles is correct. One is a version that has quietly halved the game's difficulty.

The move itself is written once, for a single direction, and the other three are done by rotating the board into that direction and back. Four near-identical loops is where almost every bug in this puzzle comes from.

Every other instrument on this bench measures an input rate, and 2048 is an input — four keys, pressed deliberately, on a board that answers each one. So the same clock is pointed at it: moves, moves per minute, and score per move.

A move only counts if it changed the board. Pushing into a wall does nothing to the grid and is not recorded, because counting it would reward holding a dead direction — the 2048 equivalent of an auto-clicker, and just as uninformative.

The rate is held back until four real moves have been made. Two moves over a fraction of a second projects to a rate nobody sustains, which is the same error a one-second click test makes when it is quoted without its interval.

Raw score rewards survival. Play carefully enough for long enough and it climbs regardless of whether you are playing well, which is why two players with the same score can have had completely different games.

Score per move divides that out. It rises only when your merges are large, because the score added by a merge is the value of the tile it creates — one 256 merge is worth thirty-two 8s. A game averaging under 8 points a move is mostly shuffling; over 20 means the big merges are landing.

It is also the figure that exposes the most common mistake. A player who panics and alternates two directions will keep their score creeping up while their score per move collapses, and the number says so several moves before the board does.

Unlike a click test, a high rate here is usually bad news. 2048 has no clock and nothing rewards haste — a fast move rate alongside a falling score per move is the signature of someone reacting to the board rather than reading it.

Strong play tends to sit between 40 and 80 moves a minute, with pauses where the board needs thinking about. Sustained rates well above that generally mean the corner discipline has gone and the moves have become reflexes.

That is the genuine difference between this page and the click tests next door. There, a higher number is a better result. Here it is a diagnostic, and the number worth improving is the one beside it.

Choose a corner and never make the move that would pull your largest tile out of it. In practice that means using only two of the four directions for most of the game, bringing in a third to reorganise, and treating the fourth as something close to a loss.

The reason it works is that a new tile appears after every move, in a random empty cell. Keeping your largest tile anchored means the spawn can never land behind it and split your descending run — and a split run is what actually ends games, rather than the board simply filling.