Identity Without the Dangerous Shortcut of “Looks Similar”
For Utopia, resolving an entity is not a cosmetic cleanup step after extraction. It decides which history a new mention joins, so this chapter follows a mention from name recall through similarity, review, merge, and reversion. The useful judgment is why Utopia treats a possible merge as a higher-risk transition than a temporary split, and how human decisions become bounded precedent without allowing the model to turn its own past guesses into authority.
A mention moves through states, not straight to a name
A document may refer to an organization by its full legal name in one passage and by a shortened name in another. Two unrelated people may share exactly the same name. A product may appear with a version suffix in one source and without it in another. The text supplies names; it does not supply a globally unique identifier. Utopia therefore begins entity resolution with recall, not with a forced answer about identity.
The first recall path looks for an exact match on a canonical name or an alias. If a previous merge moved one name into an entity’s aliases, that alias is part of the bridge between future mentions and the existing entity. A clear match can attach the mention and update the entity’s profile. If there is no direct match, Utopia uses the mention’s profile and context to assess similarity. The purpose is not to make a score look precise. It is to separate cases that can be attached safely from cases that need a second decision.
A clear similarity match attaches to an existing entity. A gray-zone match creates a new entity and places the pair in review. A dissimilar match creates a new entity without touching the review queue. This may seem conservative, but the cost of the two errors is asymmetric. An extra entity can be merged later after better evidence arrives. A wrong merge moves facts from two histories onto one node, so later extraction, reasoning, and answers can inherit a false combination.
Utopia adds containment recall because equality alone has a predictable blind spot. A shorter name may be contained in a longer name, but the shorter form cannot be enumerated as an alias in advance. A mention such as a surname or abbreviated organization name can therefore become a second entity even when the longer form already exists. Containment recall looks for that relationship under restrictions: very short generic words are not useful identity evidence, and the number of candidates admitted to review is capped so one common name cannot flood the queue. The cap is a deliberate capacity boundary, not a claim that every possible duplicate has been discovered.
Similarity is also filtered by type relationships. A name collision across clearly incompatible entity types should not be merged merely because the spelling is identical. Conversely, different labels do not automatically establish different identity when the labels belong to compatible families or when the surrounding facts support one entity. The system’s ontology supplies a signal, but it does not turn a type label into an infallible identity key. A class can be too coarse, a mention can be untyped, and a model can choose a neighboring class.
That last case is why Utopia permits duplicate names and even duplicate untyped names to coexist.