How object model edges are built: origin and method of each link, how claims resolve or stay open, what the wanted list groups, and what is never guessed.
Every edge answers three questions: what links to what (source and target are type ids), how the link was found (origin and method), and how much evidence stands behind it (links, from_items, unlinked, and one example). relation is either a claim's own relation type, such as the way one contract refers to another, or one of the structured relations evidences, names_party, keyed_to and shares_key. key names the field joined on: the canonical key field for a shared_key edge, the party field for the party_* methods.
Edge methods
relation_claimdocument_textA resolved claim: a document states a relation and the cited identifier matches a document on file. links counts the resolved claims.
claim_recorddocument_textAn open claim whose cited identifier is a golden record's id in a system of record. The edge ends at the master type. Ids that two records or domains share are dropped, so a workspace whose systems reuse number ranges loses these links rather than mislinking.
master_evidencestructuredDocuments attached to golden records by the Master Records evidence rule. links counts the attached documents.
party_namestructuredThe party field of a bound source matched by name to golden records: by exact folded name, or by the best name similarity above a configured threshold. A tie attaches to nothing.
party_crosswalkstructuredRows of a bound source keyed to golden records through the crosswalk of a system of record.
shared_keystructuredTwo bound sources carrying one canonical key field that share values. links counts the shared key values.
unlinked counts the source-side items that were checked and found no link, which is how a graph of 187 matches shows its 25 misses; it is null when the count is not meaningful for the method. truncated is true when a source was read only up to the configured row cap. origin is document_text only for the two claim methods; every join on system ids or captured values is structured, so a reader can separate what was read from prose from what was joined deterministically.
Nothing Is Guessed Into a Link
The model prefers a missing link to a wrong one. A party name or system id that fits two golden records equally links nothing. A key value shorter than three characters (four for system ids) or made only of zeros links nothing. A structured edge is built only from configuration the workspace already holds, and the model never infers a join from field names it was not told about. If a link you expect is missing, the likely cause is one of these refusals or a document hidden from the key's minting user, not a lost record.
The Wanted List
A claim is open when the identifier it cites matches no document on file. Open claims that cite a golden record's id become claim_record edges. The rest, the explicit claims with a primary key that nobody on file carries, are grouped in wanted by source type, relation and key type. Each group reports claims, the distinct documents that cite them, the number of distinct identifiers, and one example with the raw cited text, the quoted words, the document, its file name and the page. The list is a to-do list for the workspace: a document to load, or a reference to correct.
Wanted fields (wanted[])
sourcestringId of the type whose documents carry the claims.
relationstringThe claims' relation type.
key_typestringThe primary key type the claims cite, for example contract_no or order_no.
claimsintegerOpen claims in the group.
documentsintegerDocuments that carry them.
identifiersintegerDistinct cited identifiers.
exampleobject | nullOne claim: raw (the cited text), quote, document_id, filename, page.
Relation claims exist only for documents whose extraction carries identity keys and relation claims. Documents captured by a Spec run alone have none, so their types appear in the graph with structured edges but without document-text edges until their claims have been extracted.
Frequently asked questions
What does unlinked mean on an edge?+−
The number of source-side items the join checked that found no link. It is null where the count is not meaningful for the method. Together with from_items it gives the match rate of a structured edge.
Why is a link I expect missing?+−
A match that fits two golden records equally, a key value that is too short or all zeros, or a document hidden from the key's minting user links nothing, by design. A claim also needs the document's extraction to carry relation claims before it can produce a document-text edge.
What is the difference between claims_unresolved and claims_on_record?+−
Both are open claims. claims_on_record counts the ones whose cited identifier is a golden record's id in a system of record, which become claim_record edges. claims_unresolved counts the ones that cite nothing on file, the pool of the wanted list.
What does truncated mean on an edge?+−
A source behind the edge was read only up to the configured row cap. The counts on that edge are lower bounds.