Method and evidence
This page documents the public method: where the word lists come from, what is filtered during import, how matching works in the browser, and why every final candidate still needs to be checked against the rules you are using.
| Source repository | hermitdave/FrequencyWords at commit 525f9b560de45753a5ea01069454e72e9aa541c6 |
|---|---|
| Corpus named by source | OpenSubtitles2018 frequency data |
| Content licence | CC BY-SA 4.0 |
| Languages | English, Spanish, French, German, Portuguese |
| Published list size | 20,000 unique frequency-ranked source tokens per selected language |
| Correction contact | Pink Banana Tools contact |
The published data files keep the first 20,000 unique frequency-ranked tokens per language after filtering to entries that normalize to 2–18 Latin-script letters, with optional internal apostrophes or hyphens. The pinned build script rejects cross-script lookalikes before writing each file hash to the manifest. The site does not claim official dictionary status. The adapted lists are redistributed under the same CC BY-SA 4.0 terms.
Input and candidate words are lowercased and accent-normalized for matching. That lets a practical rack such as carte find écart, while the result still displays the source spelling for verification.
Each result must be buildable from the letters you typed. Repeated letters are counted, so a candidate needing two copies of a letter needs two input letters unless a wildcard covers the gap.
Each ? stands for one unknown letter. Several wildcards or a very long rack can produce broad lists, so the starts, contains, and ends filters are meant to narrow searches when the puzzle gives a real constraint.
The starts, contains, and ends fields are normalized the same way as the rack. They are applied after the letter-count check, so they reduce already-buildable candidates rather than searching unrelated words.
Matching words are grouped by normalized length, longest first. Within the same length group, results are shown alphabetically. The imported data is frequency-ranked, but the display order favors scanability for puzzle work.
The selected data/words-<lang>.json file loads on demand in the browser and is cached in memory while the page stays open. Letter matching happens locally; the tool does not need a backend or text submission.
The lists come from frequency data, not official Scrabble, Wordle, crossword, classroom, or publishing dictionaries. Slang, proper names, inflections, regional variants, and adult or offensive language may be present or absent.
Verified on July 29, 2026: the public method page, manifest, and local tests refer to 20,000 entries for each selected language, the pinned FrequencyWords commit above, OpenSubtitles2018, and CC BY-SA 4.0. Results remain candidates and should be checked against the user’s game or publishing rules.