Entity resolution

Records arrive from fifty Secretary of State registries, licensing boards, permit and inspection files, government datasets, web pages, and card transactions. None of them carry a common key. Names are spelled differently, addresses are formatted differently, and the same suite number appears five ways in four files.

Resolution is the step that decides which of those records point at the same real thing. It happens before any question you actually care about — coverage, monitoring, revenue, ownership — because every one of those answers is a consequence of where the boundaries were drawn.

Get it wrong in one direction and two businesses are fused into one profile, so you clear an applicant on somebody else's evidence. Get it wrong in the other and one business splits into six thin profiles, so a real applicant looks like a stranger. Both errors are expensive, and they pull in opposite directions.

A brand is the name customers know. It is what people type into a form. One brand can sit above many legal entities, and one legal entity can carry several brands.

A legal entity is the thing that files. It exists in a state registry, has a status, a formation date, officers, and a registered agent. It can be dormant, or exist purely to hold a lease.

An operating location is an address where work happens. It has hours, a phone number, reviews, inspections, and card transactions. Locations open and close on a schedule that has nothing to do with the filing calendar.

These form a graph, not a tidy hierarchy. Enigma indexes the three with colour throughout this site — brand, legal entity, location — because which layer you are looking at usually explains the disagreement in front of you.

Deterministic rules — exact name plus exact address — break on the first “Ste 200” versus “Suite #200”. They break again on a legal name that changed in 2019 and a brand that never did.

So a matcher scores evidence instead: name and token similarity, normalized and geocoded addresses, officer and agent overlap, shared phone numbers and domains, and time — a registration that lapses in the same quarter a location goes dark is a different signal than one that lapses alone.

Every threshold is a trade, and the trade is not symmetric across use cases. An onboarding flow that wants coverage tolerates a looser merge than a sanctions program that must not attach the wrong person to the wrong company. That is why resolution is published with a precision figure rather than a claim of correctness.

The graph holds 100M+ registrations, 30M+ brands, and 2B+ relationships, resolved at 99.2% pairwise precision.

Pairwise precision is measured by sampling pairs of records the system says belong to the same business and checking how many actually do. It is the right measure to publish, and the wrong one to read as “99.2% of profiles are perfect.” A cluster of 100 records contains 4,950 pairs; a small number of bad pairs can still sit inside a single profile you are reading.

The second limit is more important than the first: a filing is not a business. Resolution can link forty registrations to one entity with high confidence and still be describing paperwork. Whether anything operates is a separate question, answered by evidence tiers, not by the match.

Query a brand and the answer is a shape, not a row. This is the Blue Bottle Coffee fixture — the graph the illustration below draws, and the same records behind the demos elsewhere on this site.

One brand resolves to two legal entities. Those entities carry forty registrations across twenty-nine states, and beneath them sit ninety-nine operating locations with $64.0M in observed card revenue.

The useful part is what a narrower query would have returned. Match only the legal name in California and you see one filing, formed in 2008, with no sense of whether it is a coffee company or a holding shell. The shape is the evidence.