Same object and registry
Compare identifier system and value together, not an isolated character string. An establishment and its legal unit need linked records.
OSINT / IDENTITY
Match entities without merging namesakes
Two similar records do not necessarily describe the same object. Matching must retain sources, dates and decision reasons, including when no merge is possible.
Choose the unit you need to identify
First state whether the record represents a person, legal unit, establishment, brand or research organisation. A relationship between objects is not identity. A normalised name helps find candidates; it is not a merge key.
Identifiers do not all identify the same object
| ID | Identified object | Check |
|---|---|---|
| SIREN | Legal unit | Nine digits. Check the registry and store as text to preserve leading zeros. A trading name does not replace the identifier. |
| SIRET | Establishment | Fourteen digits; the first nine are the SIREN. Two SIRETs with the same prefix may refer to one legal unit but to two distinct establishments. |
| ORCID iD | Person in research | Retain the full URI and leading zeros. The last character can be X. An ORCID identifies a person record; it does not automatically certify every profile statement. |
| ROR ID | Research organisation | Review the candidate, location and relationships. An affiliation string may mix department, parent organisation and address; do not automatically pick the first similar name. |
SIRENLegal unit
SIRET AEstablishment
SIRET BEstablishment
Five checks before merging
Source for each attribute
Retain the URL or document, observation date and known effective date for names, addresses and status. An aggregated profile does not replace the source document.
Compatible period
Different addresses may reflect a move; another name may be historical. Check periods before replacing an old value.
Visible contradictions
Different IDs in the same registry or incompatible places require review. Do not erase conflicts to tidy the record.
Reversible merge
Keep original row IDs, rule, reviewer, date and a mapping table. Later corrections must allow records to be separated.
A repeatable method
Keep raw data
Store original labels and identifiers separately from normalised values.
Generate candidates
Normalise spacing and case for searching. Preserve accents, alternative names and transliterations as variants; avoid dropping suffixes that distinguish objects.
Compare independent records
Check the ID in its registry and other attributes against appropriate sources. Two portals copying one registry are not independent confirmations.
Classify the decision
Use “documented same object”, “distinct objects”, “relationship to retain” or “undetermined”. Similarity is not certainty.
Review the effect on the dataset
Compare row counts, broken links and lost attributes. Review rejected matches too, to find missed duplicates.
Four common decisions
These are schematic cases, not accounts of real people or businesses.
Same name, different identifiers
Decision : Keep two candidates
Decisive check : Check registry, object and period; a name alone cannot decide.
One SIREN, two SIRETs
Decision : Link establishments to the legal unit
Decisive check : Keep establishment records, addresses and dates. Do not merge them into one location.
Same ID, old address
Decision : Retain the history
Decisive check : Check source and effective date; distinguish a real change from a typing error.
Department and parent organisation
Decision : Create a documented relationship
Decisive check : Check record granularity and institutional relationship. Affiliation is not identity.
Record to retain
For each decision retain: record IDs A and B; object type; registry and identifiers; raw labels; sources and dates; agreements; contradictions; reasoned decision; reviewer; revision date. A missing field means “not supplied”, not “false”.
One row per decision, retaining both original records, their sources, conflicts and reasons. The template is blank; use “undetermined” when the evidence cannot settle the match.
Download the matching decision register CSVPractical questions
Does a valid identifier format prove existence?
No. Format or checksum catches some typing errors. Check the official registry record and whether it identifies the intended object.
Can I merge on an address or web domain?
Not automatically. Multiple organisations may share an address and one website may represent a group. Treat these fields as clues.
How should translated names be handled?
Retain published name, language, variants and a verified shared identifier. Similar transliterations do not prove two people are the same.
Do different IDs prove objects have no relationship?
No. They may identify distinct related objects or belong to different registries. Document the relationship instead of inventing equivalence.
Should all matched attributes be published?
No. Collection, internal analysis and publication are separate decisions. Keep attributes needed for the subject and adapt disclosure.
How is automatic matching evaluated?
Use reviewed cases with false matches and missed duplicates. Measure precision and recall within the test set’s stated scope; a similarity score is not a reliability rate.