FND001 · Network Intelligence

Entity Resolution

FND001 Entity Resolution is the foundation capability that resolves duplicate representations of one real-world entity into a single canonical identity, so that risk attaching to any representation attaches to all of them.

Without it, an entity appearing as Alpha Trading Ltd, ALPHA TRADING LIMITED and Alpha Trading is three parties carrying a third of the risk each, and no typology can see the aggregate. Every network product downstream is a statement about edges between entities, and is only as sound as the identity of the nodes.

A typology asks whether behaviour is suspicious. FND001 asks the question that must already be answered before that one means anything: are these records the same party?

Risk indicators

Risk

• Unresolved duplicate understating exposure across every downstream product
• Name gate cleared but the match carried thereafter by address alone
• Name gate cleared but the match carried thereafter by telephone alone
• Names align while registration numbers differ
• Address shared by many otherwise dissimilar names
• Merge uniting entities incorporated in different registries
• Match score near a confidence band boundary

Investigation actions

1. Read the matched attributes rather than the score alone
2. Confirm the name gate was cleared before relying on a supporting attribute
3. Treat a match supported only by a shared address as a co-location finding
4. Re-read every typology finding against both representations once a match is accepted
5. Record the reason whenever a match is rejected
6. Treat an unresolved pair as a standing understatement of exposure, not a closed question

Detection logic

FND001 scores a candidate pair by weighted attribute agreement. The weights are governed constants in the entity resolution engine:

Name similarity 0.40

Registration number 0.30

Address 0.15

Email address 0.10

Telephone 0.05

The name gate is the load-bearing rule. Registration, address, email and telephone contribute only when name similarity already exceeds 0.85. A shared address between two dissimilar names is not evidence they are the same party. It is evidence of a shared address, which is a different product entirely — PAT005 Common Address. Removing that gate would collapse the two capabilities into one and make every serviced-office tenant a duplicate of every other.

The score is capped at 1.00 and rounded to four decimal places.

Name similarity is token-aligned and suffix-insensitive. Each token of one name is aligned to its best Jaro-Winkler counterpart in the other, the comparison is run in both directions, and the lower of the two totals is normalised by the longer token count. Taking the minimum is what makes the measure symmetric: a short name contained inside a longer one cannot score as a strong match on the strength of the generous direction alone.

The consequence is deliberate and worth stating, because it looks like a miss. Gamma Imports Ltd and Gamma Imports Limited resolve at 1.00, since corporate suffixes are stripped before comparison and a typo is tolerated. But Omega Trading and Omega Oil Trading fall below the gate, and so do Theta Investments and Theta Investment Holdings. An extra distinguishing word is penalised by design — it is usually a different company, not a spelling of the same one.

Curated from fc-03-network-intelligence-library/intelligence-products/specific/FND001/README.md