Word Hunt

Press a letter and drag through the ones beside it. Any of the eight neighbours counts — across, down and both diagonals — and you may not land on a cell you have already used in that word. Let go at the end of a drag and the trail is submitted: if the letters spell a word of three letters or more, it scores and joins your list. You can also tap the letters one at a time without dragging at all, which is easier on a small screen — the trail stays where it is between taps, and tapping the last letter a second time takes the word. Either way the clock is the whole game; when it runs out the round closes and every word you did not find is listed for you, longest first.

The board tells you what it thinks as you drag. The trail you have drawn so far is lit, the letters you have picked up appear above the grid, and the moment the trail spells something the dictionary knows, the read-out turns green and shows what it is worth. Nothing is committed until you release, so a trail that has wandered somewhere useless can simply be backed up.

A word is worth the square of its length minus two. Three letters is 1 point, four is 4, five is 9, six is 16, seven is 25, eight is 36. It is steeper than it first looks, and it is steep on purpose: a 4 × 4 grid contains 408 paths of exactly three cells and 6,712 of five, so a five-letter word is not twice the work of a three-letter one, it is sixteen times the haystack. The scoring says so.

Each round has a par — a score the round is asking for, set from that particular grid rather than from a flat number, because grids differ wildly in how much they hold. Sprint asks for 4% of everything its grid contains, Classic for 8% and Big grid for 7%, each clamped so a freakishly rich grid does not set an impossible target and a thin one does not hand you a free pass. Reach par and the round counts as a win in your statistics; fall short and it counts as a loss, which is what makes the win rate mean anything.

At least 60 on a 4 × 4 and at least 140 on a 5 × 5, and that is a guarantee rather than an average. Before a grid is shown it is completely solved — every one of the 172,823 words in the validation list is tested against it — and if the count comes in under the floor the grid is thrown away and rebuilt from a fresh sub-seed. Typical grids hold far more than the floor: measured over a sample of 120 dealt grids, the median 4 × 4 holds 127 findable words and the median 5 × 5 holds 258.

The floor is not the only guarantee. A real 5-, 6- or 7-letter word is planted first, laid along a winding path of touching cells before a single filler letter is placed, so there is always at least one long word in there to go looking for. It is drawn from the 25,322-word familiar list, never from the obscure tail — a grid whose only long word is zuurveldt is not a hard grid, it is a broken one.

• Plant — a familiar 5-to-7-letter word is laid along a self-avoiding path of adjacent cells.
• Fill — every remaining cell is drawn from the weighted letter bag.
• Repair — the vowel count is pushed inside three-in-ten and five-in-ten, without touching the planted cells.
• Solve — the whole validation list is traced against the grid.
• Reject — under the floor, the grid is discarded and rebuilt from the next sub-seed.

The letters come from a bag of 101 tiles whose weights are the measured letter frequency of the shipped familiar word list, counted across its 3-to-9-letter entries because those are the lengths a grid actually yields. E gets 12 tiles, S and A get 8, J and Z and X get 1 each. Of the 101, 36 are vowels, and the grid is then repaired so that no fewer than three cells in ten and no more than five in ten are vowels. Both bounds matter: a wall of consonants is unplayable, and so is a puddle of vowels.

Q is left out of the bag altogether, and a word containing one is never planted either, so no grid here can show a Q. Its frequency across the familiar list is 0.16% and rounds to zero tiles, and a Q is dead weight unless a U lands next to it: only 29 of the 2,527 validation-list words containing a Q — one in 87 — have a Q that is not followed by a U, and the bag holds just 3 U tiles in 101. A cell that needs one specific neighbour before it means anything is a hole in the board, not a letter, and sixteen cells is too few to spend one that way.

Yes, and this is the rule implementations most often get wrong. The only restriction on a path is that no cell is used twice in the same word. The shape it draws is irrelevant: it may cross its own earlier leg diagonally, run back alongside itself, or spiral into the middle of the grid. A board that rejects a self-crossing trail has invented a rule, and the player who traced a perfectly good word and was told it was not one does not come back to find out why.

Cells are also only locked for the word you are drawing. Finding tea does not use up the T for the rest of the round — the same four cells might go on to make team, mate and meat. A 4 × 4 grid contains 12,029,540 traceable paths of three cells or more, which is why a grid that holds 126 words still feels like it is hiding them.