One Word Map, Five Games: How Semantic Similarity Became Proximity, Traverse, Quartet, Outlier and Huddle

The five word games on this site look like five different ideas. One is a daily guessing game, one is a chain, two are sorting puzzles, one is something you pass around a table. Underneath, they are the same idea applied five times: treat a word as a point in space rather than a string of letters, and let the distance between two points stand for how close they are in meaning. This article is about that map — what it is, how each game uses it, why it ships inside the app rather than living on a server, and the things it gets wrong.

The map

A word-embedding model is a table that assigns every word in a vocabulary a list of a few hundred numbers. Those numbers are not chosen by hand. They are learned by reading a very large body of English text and adjusting each word's numbers so that words which appear in similar contexts end up with similar lists. After training, the list for tea sits near the list for coffee, both sit near cup, and all three sit a long way from bicycle. Nobody told the model that tea and coffee are drinks. It inferred it from the company they keep.

Once you have the table, “how close are these two words” is arithmetic: compare the two lists and get a number between roughly minus one and one. The companion article goes into what that number means and why the games convert it into a rank. For now the important thing is that the map is a fixed, finished object. It can be computed once, packaged, and shipped.

Game one: navigate to a hidden point

Proximity is the most direct use of the map. Pick a secret word. For every word in the vocabulary, compute its similarity to the secret, and sort. The secret is rank 1; the next-closest word is rank 2; and so on out to the far edge of the vocabulary. A player's guess is scored by looking up its rank in that sorted list. That is the entire scoring function. There is nothing about letters, length or spelling anywhere in it, which is why guessing tea can score 3 while guessing teal scores 700.

The daily word is the same for everyone because the sort is deterministic: same secret, same map, same ranks. Practice mode picks a random secret and re-sorts. The hint that reveals “a word roughly halfway to the answer” is a look-up too — take the player's best rank so far, halve it, and return the word at that position.

Game two: walk across the map

Traverse uses the same similarity but adds a rule that turns the map into a graph. Every word is connected to its twelve nearest neighbours and to nothing else. A chain is valid only if each word is connected to the one before it. The start and end word are chosen, the shortest path between them is found with an ordinary breadth-first search, and its length becomes par.

Twelve is not a round number picked for elegance. With fifty neighbours per word, almost every pair of common words is two or three steps apart and the game has no room to be interesting. With five, rare words become dead ends — all five of their neighbours point back toward the common word they came from. Twelve was where crossings needed several steps but nobody got stuck. Every bridge is verified to have a path before it is allowed into the game, which is a stronger guarantee than a hand-written puzzle can make.

The “#19 nearest” message when a word is rejected is the map again: the rejected word is not in the top twelve, but the game tells you where it did fall, so you learn whether you were close to a valid step or nowhere near one.

Games three and four: check a puzzle before anyone plays it

Quartet and Outlier are category games, and categories are not something an embedding model gives you directly. Both games use a separate, hand-built database of words tagged with categories — mammals, currencies, kitchen things, weather. The map's role is different here: it measures how far apart two categories are, which is what difficulty means in these games. A hammer among fruit is easy because tools and fruit are distant regions of the map; a crab among fish is hard because they are neighbours.

The map also helps hunt for the failure that ruins puzzle games of this kind: the second answer. If any word in a Quartet grid sits almost as close to a second group as to its intended one, or if three of Outlier's four “belonging” words form a tighter cluster than the four do together, the puzzle has an argument in it, and arguments are not fun. The checking process is worth its own article, but the short version is that a round is only allowed to ship if it has exactly one answer, and the map is one of the tools used to prove it.

Game five: write cards nobody had to write

Huddle is a party game, and the map is doing something quieter there. In Forbidden, the five banned words on a card are the five words nearest in meaning to the target. Nobody sat down and decided that a card for ocean should ban sea, water, wave, beach and deep; the map produced that list, and it is the list a human would have produced on a good day, every time, for hundreds of cards. In Impostor, the impostor's word is a near neighbour of the real one — close enough that a one-word clue can plausibly come from either, far enough that a careful table can catch the difference.

Hand-written decks for games like this are inconsistent. Some cards are impossible, some are trivial, and the person who wrote them knows the answers. Generating the decks from the map gives every card the same amount of difficulty by construction.

Why the map is inside the app

The obvious architecture is a server: the app sends your guess, the server looks up the rank, the app shows it. Every game here does the opposite. The vocabulary and the map are packaged in the app, and the look-up runs on your phone.

Partly that is a privacy decision. There is nothing to send, so nothing is sent; your guesses, your streaks and your statistics do not exist anywhere except on your device. Partly it is practical: the games work on a plane, on the underground, and anywhere else a connection is poor, and they never break because a server is down. The cost is app size. A few hundred numbers per word across a large vocabulary adds up, and each app spends effort trimming the vocabulary and compressing the numbers to keep the install small. The only thing in any of these apps that needs a connection is the advertising that keeps them free.

What the map is bad at

It is worth being honest about the failure modes, because players hit them.

• Opposites are close. Hot and cold appear in the same sentences — “the water was hot,” “the water was cold” — so the model places them near each other. In Proximity that means guessing the opposite of the answer scores well. This is a property of how the map is learned, not a bug that can be fixed without a different kind of model.
• One spelling, several meanings. Bank has one point on the map, sitting somewhere between the river and the money. Words like that behave unpredictably in every game, and the puzzle games avoid them where they can.
• Rare words have poor neighbourhoods. A word the model saw only a few hundred times during training has a fuzzy position. Traverse's dead-end problem is mostly rare words; the vocabulary is trimmed to reduce it.
• The map knows usage, not definitions. Coffee and morning can be closer than coffee and beverage. Players who think in dictionary terms find this frustrating at first and useful once they adjust.

None of this is hidden. Proximity's rank tells you precisely how the map sees your guess; Traverse tells you where a rejected word fell; Outlier shows you the category you missed. The games are built to expose the map rather than pretend it is an oracle, because the map is the game.