The full dataset is not published yet. A license review comes first, and the dataset will be released once it is finished; no date is promised. Until then the site shows the dataset's statistics and a sample, and lists cannot be built or downloaded.

Common English OpenList
Sections

How it was built

From the English OpenList's 9,653,999card:eol_universe_rows strings to the 10,020,150card:rows_universe rows of version 1.0, in 7 steps. Each step has a committed record, and the counts are those records'.

  1. The English OpenList's universe

    The EOL's valid and invalid lists at revision c1139698: every string the EOL had taken in, whichever way it ruled on it.

    Record: reference/research/merge/merge.json

    9,653,999card:eol_universe_rows rows

  2. The census

    The pinned sources read again under the spelling grammar, keeping the candidates in the classes the EOL's letters-only rule could not hold: hyphens, accents, a diaeresis, apostrophes, ligatures. Merged with the EOL's universe, 1 row per string; filtering the merge by the EOL's flag gives back its pinned lists byte for byte.

    Record: reference/research/census/census.json

    +213,867card:census_candidates

    9,867,866card:merge_rows rows

  3. The audit

    Every source's coverage of the merged universe, reconciled to a remainder of 0: what a source listed and the merge lacked was added. The result is the single-token pin.

    Record: reference/research/universe/universe.json

    +133,875card:c4_additions

    10,001,741card:v1_rows rows

  4. Multiword terms

    WordNet's and Wiktionary's multiword entries, admitted under the grammar's space rule. The pin's single-token part reproduces the earlier pin byte for byte.

    Record: pipeline/research/universe_v2/universe_v2.json

    +294,865card:pin_spaced

    10,296,606card:rows_pin rows

  5. The layers

    No row is added. Each row is measured against the 13 panel lists, given its features and its record of how it came in, and marked where the EOL's verification pipeline or a model alone decided it, or its synthetic engine made it.

    Record: pipeline/research/provenance/provenance.json

    no row added or removed

    10,296,606card:rows_pin rows

  6. Restricted rows

    The rows that only restricted lists attest are in no version.

    Record: pipeline/research/release/release_v1.0.json

    −1,985card:rows_restricted

    10,294,621card:rows_shippable rows

  7. The public projection

    The withheld lists, labels, and feature columns removed, and the rows left with no public evidence held back in the private record. What remains is version 1.0.

    Record: pipeline/research/release/release_v1.0.json

    −274,471card:held_back_rows

    10,020,150card:rows_universe rows

The release build then asserted, on each of its 7card:files_asserted files, that none of the 1,985card:restricted_words_checked restricted rows is present, that every id is distinct, that each config is its rule's filter, and that no withheld column, label, or held-back row is present. The bars above share 1 scale, the largest change, so the changes compare with each other; every total is about 10 million.

Language models and this dataset

No language model decided any row the CEOL added: the census, the audit, and the spaced admission follow the grammar and the pinned sources, and every figure on this card is computed by committed code from committed records. The feature layer's character model (§6) is an interpolated character n-gram model, not a language model.

The EOL decided some rows of its valid list with a language model. Its valid dictionary names the model on each of those entries: Google Gemini 3 Flash Preview, through its verification pipeline or by the model's verdict alone. The EOL's own report says the earlier phases of its pipeline used Gemini 2.0 Flash and Gemini 2.5 Flash. model_or_engine marks these rows (§1, §5), so a list can keep them or leave them out.