DEMO-2026-0042

Entity screening uses an exact-first algorithm: exact identifiers/documents are evaluated first, then exact normalized name-token matches, edit-distance candidates, fuzzy name matching using normalized aliases, token-sort values and phonetic keys, and optional address similarity evidence when an address was supplied. The request threshold was 80 and the list filter was all active sources.

The phonetic key reduces a name to its consonant skeleton. This is what allows a name written in one script to match a list entry written in another — Cyrillic and Latin spellings of the same company share no character sequences at all. The same reduction can also connect unrelated names (the code is only 4 characters long, so unrelated surnames occasionally collide), so each fuzzy candidate additionally records how far the two spellings agree once both are transliterated into a single script. Candidates carried by the phonetic key alone are marked "sound only". For a single-word query that mark is informational only — a one-word name is the sole way to bridge two scripts, so the score is left as-is. For a query of two or more words, a "sound only" match with no matching word and no spelling agreement is capped at 60 before bonuses, so a coincidental sound match on an unrelated name no longer reaches the same confidence as a real one; a matching date of birth or country can still raise it.