Its rule and its record

What may be on the list, how a word gets on or off it, and every correction made by hand, counted from the ledgers that recorded them.

The rule

An entry is lowercase a to z and nothing else: the pattern ^[a-z]+$, with no length bound. No capital, space, hyphen, apostrophe, digit, or accent.

So the list is deliberately not a tournament word list. It holds 2 one-letter words and 3 words longer than 45 letters, the longest 47 letters long; tournament lists exclude both. The download "Lengths 2 to 45" is the list in that shape.

How a word gets on or off

The nightly run is the only automated process that adds words. It looks for new words (Merriam-Webster's and Wordnik's words of the day, and a file of manual additions) and checks about 1,000 entries of the invalid list against dictionary APIs. A word a dictionary recognizes is promoted to the list, except a single letter other than a and i: a dictionary's entry for a letter names the letter, so the run passes over them. A word whose check was refused or rate-limited is not checked that night, and nothing records it as rejected.
A correction ledger is the only other way a word moves. Each row of a ledger in corrections/ names the word, the action, the reason, the method, the source that ruled, and the date, before any list changes.
No word is invalid for good. A word moved off the list is queued to be checked again, because "no dictionary we could reach knew it that day" is a statement about the dictionaries, not about English.

What decided each word

The list's record of each word says how it got on. This site sorts the records into 4 values by 1 rule, the rule the CEOL Universe uses, so both count the same words:

Value	What it means	Words
From a source	It came from a word list, a dictionary, or a glossary of a field such as medicine. No language model or generator decided it.	176,222
Checked with a model	It passed a check of several steps, and a language model was 1 of the steps.	131,226
A model's call alone	A language model judged it a word, and no other check backs that.	6,471
Generated by rule	The list's synthetic engine made it by rule, for example by adding a prefix or a suffix to a word already on the list.	31,243

The synthetic engine made 31,243 words by rule: affixes on the list's own words, chemical names by substitution, medical and scientific prefixes, and Greek plurals. They are kept, and marked, so a project can leave them out: the download "Without the synthetic engine" does.

20,048 words carry a note from 1 pass of a language model that called them invalid, and the list kept them. The note is an opinion, not a verdict; the search can leave those words out.

The list was assembled in 3 large loads: December 17, 2025 (131,226 words), January 10, 2026 (175,935 words), and January 11, 2026 (31,243 words). Together they hold 338,404 of its words; 6,671 carry no date. Sorting by date added is mostly sorting by load.

The record of corrections

Every word moved by hand is a row of a ledger. These are their counts, read from the ledgers when the site is built.

August 13, 2026
150 words that were on both lists were removed from the invalid list. Each already carried a dictionary's ruling in its own record. The list itself did not change.
August 14, 2026
33,594 synthetic forms were moved to the invalid list: 16,476 comparatives and superlatives, and 17,118 gerund plurals, verb forms, and agent plurals. They were decided by the inflections a dictionary records for their stems, generalized from a sample. Every one is queued for another check.
August 16, 2026
The rule changed to ^[a-z]+$ with no length bound. 190 entries that broke it (hyphenated or accented) were deleted; 8 single letters were moved to the invalid list; and 37 strings that were on neither list were added to the invalid list as candidates, never to the list.
August 17, 2026
8 single letters were queued to be checked again under the new rule.
From August 18, 2026
The nightly run added 63 words, the last on October 30, 2026.
October 31, 2026
4 single letters the nightly run had promoted, before it passed over letters, were moved back to the invalid list. Like the other single letters, they are queued to be checked again.

The list was last published on November 4, 2026.

The ledgers also hold 12,437 rows of samples that were checked and moved nothing; they are the evidence for the corrections above. How each was decided is in corrections/README.md.