AI proposes. A chain of checks decides.

Verbobum is published by Alexandre Désir. An AI model first proposes an original definition and examples for a chosen language and topic. The cards are not presented as manually written by a lexicographer.

Every decision keeps an audit trail. If a check cannot finish, a source is missing, or reviewers disagree, the candidate is rejected.

Eight first-stage checks cover structure, attested existence and part of speech, inflected forms in examples, length, grammar, selected proper names, distance from source wording, and usage register.

Two models from different providers independently match the candidate to an attested sense and judge its examples. Both must select the same sense and approve the examples; there is no tie-breaker that forces publication.

Kaikki’s open Wiktionary extracts verify existence, part of speech, forms, senses, and register tags. LanguageTool checks example grammar. Source wording is used for verification, never copied into the published definition.

For historical articles, an agent selects only words that appear to have a documentable journey. Web research is restricted to approved linguistic and heritage institutions; every source must respond, every factual claim must retain a citation, and the evidence file needs at least three pages across multiple domains.

A writer then adapts the story for all six languages without adding facts. A deterministic gate checks structure, sources, and citations, and a model from another provider looks for unsupported claims and mechanical translations. Even after those checks, the result remains a draft: a person must approve it and then trigger publication separately.

Automation makes regular production across six languages possible; strict checks prevent speed from becoming permission to publish anything. The method reduces risk but cannot promise perfection. Report a suspected error to daily@verbobum.com with the word and language.