One people, one page: identity matching at catalog scale

Under the hood

• Identity by article: the review resolves every candidate to a Wikipedia article. Two names for one people resolve to one article, so normalizing the article title and matching it against the live catalog turns an identity question into a lookup. That is how Ngbaka meets Mbaka, a pair that string distance alone cannot tell apart from noise.
• String matching runs alongside it: small edit distances, and a sound-folded consonant skeleton that folds doubled letters and interchangeable consonants such as c, k and q.
• Wave 3 drew on 476 well-documented names, each containing the name of a live page inside its own. Read one by one, many were coincidences of spelling: Rashaida contains Haida, Peranakan contains Akan, Miami contains Ami. The Pontic Greeks passed as a distinct subgroup of the Greeks, whose page was already live.
• Passing candidates then clear a five-way screen: the same Wikipedia title as a live page, a Wikipedia redirect, the title's slug already live, a normalized-name twin, and an exact slug.
• Any-language grounding: when no English article exists, the review sheet falls back through a set list of Wikipedia editions (French, Spanish, Portuguese, Russian, Indonesian, German and on) and records the language and title. The rubric review judges that lead as it would an English one, and the author reads it in its language and writes in US English. At the time, 578 candidates in the remaining pool with no English page had one in a major edition.