The authoring engine: 10,837 entities, one rubric, 253 new pages

Under the hood

• The harvest feeds four stages: review sheet, rubric review, authoring and publish. Candidates run in order of Wikidata sitelink count, so the best-documented peoples publish first, and every page enters the catalog hidden and goes public with a single switch.
• Duplicate detection normalizes names after stripping a trailing group noun, so "Zulu people" meets "Zulu". A containment check flags pairs such as Baloch and Balochi, or San and Khoisan-San, for review instead of rejecting them, because the same test also fires on distinct neighbors such as Han and Hani.
• The rubric has nine classes, tested in a set order. Every candidate leaves with exactly one outcome and one reason code, and the review sheet shows its Wikidata ID, its sitelink count, the opening sentence of its Wikipedia article and any live pages it resembles.
• A language model reads the Wikipedia lead and authors each page: a description of 351 to 457 words in wave 1, plus structured fields for homeland, homeland region from a 23-region vocabulary, language, religion, subgroup, phenotype profile and portrait style register. A page is written when it passes validation, and nothing is invented to fill a gap.
• Many historical peoples, such as the Normans and the Xiongnu, carry no era word in their names, and Wikidata marks few of them: 22 entities in the whole candidate pool carry a dissolution or end date. The rubric's test is whether any living community identifies by the name today, which places the Beothuk and the Batavi in the historical cohort and keeps the Hopi and the Karelians among the living.