Where the answers come from, and what can never be one
5 min read
Every daily word game eventually has the same bad morning: the answer is a word, it is in the dictionary, and nobody has ever said it out loud. Players do not experience that as a challenge. They experience it as the game cheating.
So the answer pool here is not the dictionary. It is a filtered subset — roughly 7% of the legal five-letter words survive — and this is what the filter does.
Two lists, doing different jobs
| Board | Words you may guess | Words that can be the answer | Survives |
|---|---|---|---|
| Easy (4 letters) | 3,903 | 600 | 15.4% |
| Regular (5 letters) | 8,636 | 646 | 7.5% |
| Hard (6 letters) | 15,232 | 1,117 | 7.3% |
The guess list is deliberately enormous, because a game that rejects a real word you typed is infuriating in a way that nothing else in the design can make up for. The answer list is deliberately small, because the answer is the one word you are entitled to recognise afterwards.
Frequency, measured twice
A word qualifies only if it is common enough — and “common” is measured against two corpora rather than one, because each is wrong in a different direction.
- A web corpus over-ranks internet vocabulary. Words like SITE and PAGE look far more central than they are in speech.
- A subtitle corpus is the reverse: it captures how people actually talk, and under-ranks anything written more often than spoken.
A word’s familiarity score is the geometric mean of its rank in both, so it has to place in each. Something that is huge on the web and absent from speech does not clear it, and neither does the reverse. Each board then applies a ceiling:
| Board | Familiarity ceiling | Why |
|---|---|---|
| Easy | 12,000 | The strictest, because four letters give you the least to work with |
| Regular | 16,000 | The standard board |
| Hard | 20,000 | Loosest, and still far short of the whole dictionary |
What can never be the answer
Frequency alone is not enough — plenty of common strings are terrible answers. A word also has to be known to WordNet as a base-form lemma with a usable, non-proper-noun definition, and that single requirement does most of the work:
| Excluded | Example | Reason |
|---|---|---|
| Proper nouns | PARIS | Not a word everyone shares, and unfair across regions |
| Plurals of words already in the pool | CARDS | Doubles the pool without adding a puzzle |
| Participles and inflections | WALKED | Same word, wearing a suffix |
| British spellings with an American twin | CENTRE | Both are real; serving either is a coin toss |
| Slurs | — | A public blocklist, applied to the answer pool |
And one filter that is not about words at all
A word also has to have at least one definition that reads neutrally, because the first hint on every board is the definition. LOST is a perfectly ordinary word, and one of its real dictionary glosses is about people who are about to die. That is a fine entry in a reference work and a terrible thing to put on somebody’s phone over breakfast.
So a word survives only if some gloss is usable as a clue. The line is drawn at glosses that are unpleasant out of nowhere, not at every mention of anything difficult.
The pools came in under target, and that was the right outcome
The five-letter board was specified to hold 1,500 answers. It holds 646. The six-letter board was specified for 2,500 and holds 1117.
Both fell well short, and the filters were not loosened to hit the number. At 646 answers a daily board runs for 1 years before anything repeats, which is longer than the question is worth worrying about. Relaxing the frequency ceiling to reach a round number would have bought nothing except worse mornings.
Where the data comes from
Everything above is built from four public sources — ENABLE for the guess lists, Princeton WordNet for definitions and the base-form test, a web-corpus unigram list and an OpenSubtitles frequency list for the two halves of the familiarity score, and a public blocklist. Licences and links are on the about page, and the build is a single script in the repository rather than a hand-curated file, so every rule above is one you can read.
Which is the point. A hand-picked answer list is a set of opinions; a generated one is a set of rules you can argue with.