The original English word data (preserved in data/original-dictionary.js) is derived from the English list in
lorenbrichter/Words:
https://github.com/lorenbrichter/Words/blob/master/Words/en.txt

The source repository dedicates its word lists to the public domain under
CC0 1.0 Universal:
https://creativecommons.org/publicdomain/zero/1.0/

Original dictionary entries are filtered to lowercase alphabetic words of 4-9 letters.
Words may contain repeated letters when the wheel has enough matching slots.
The upstream list aims to exclude proper nouns, but notes that a few acronyms
and abbreviations may remain.

To favor familiar words without excluding reasonable answers like "pear" and
"pore", the original list was further limited to words that appear in the top
50,000 entries of the English 2018 frequency list from
hermitdave/FrequencyWords:
https://github.com/hermitdave/FrequencyWords/blob/master/content/2018/en/en_50k.txt

The FrequencyWords code is MIT licensed. Its frequency-list content is
CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), attributed
to Hermit Dave and derived from the OpenSubtitles 2018 corpus. The original
dictionary is the intersection of that frequency list's top 50,000 entries
and the CC0 word list above, further limited to lowercase alphabetic words
of 4-9 letters. The resulting dictionary data is distributed under CC BY-SA
4.0. This cutoff restores ordinary words that the stricter top-10,000 cutoff
excluded, while filtering out the previously identified variants that do not
appear in the top 50,000.
October 2026 gameplay revision:
The active dictionary combines an editorial subset of the original data with
checked inflections. Sources are data/playable-words.txt and data/inflections.json;
dictionary.js is generated.
The first shortlist used the first 8,000 entries in the original ordered
list, removed names, subtitle artefacts and other unsuitable entries, then
restored familiar short words and a few nine-letter words from further down
the original list. data/excluded-words.txt records explicit exclusions.
This is a practical initial selection, not a validated familiarity rating.
The derived dictionary retains the CC BY-SA 4.0 attribution above.

Puzzles use a 9-letter answer as the source of their nine wheel slots.
Every published catalogue entry has exactly one accepted 9-letter answer
and between 20 and 40 accepted answers. There is no vowel-count rule.
Less familiar words outside the selected list are not accepted, including
rahui, huer and quai. The original data is retained for a potential future
bonus-word feature and is not loaded by the browser.

The source word list and frequency list have not been re-downloaded in this
revision. The preserved original bundle is the input snapshot.

October 23 word-family correction:
Candidate plurals and verb forms are generated from familiar bases, with source
attestation or explicit editorial approval, and verified using the locally
installed Vim English spelling data (en_gb and en_us). The snapshot records the
spelling-file hash and base families; no whole system spelling file is bundled.
Vim spelling documentation: https://vimhelp.org/spell.txt.html
Spelling verification does not establish familiarity or grammatical context.
A dated exception preserves the October 23 playtest wheel with 41 answers after
ordinary missing words were restored. Regular catalogue wheels retain 20–40.

October 27 noun-plural policy (supersedes plural acceptance above):
Noun-only plurals are excluded from the combined candidate list. Verb forms and
independent non-plural meanings remain eligible. The October 23 exception is retired.
The classification snapshot uses Princeton WordNet 3.0, Copyright 2006 Princeton
University, with editorial overrides. Source and file hashes are recorded in
data/plural-decisions.json. Full WordNet license: wordnet-license.txt
Source: https://wordnetcode.princeton.edu/3.0/WordNet-3.0.tar.gz
Morphology: https://wordnet.princeton.edu/documentation/morphy7wn
No WordNet words are added to the playable list; its data only informs filtering.