On this page
Record metadata reference
Look up the documented metadata keys returned with Golden records, history, quality, operations, and search results.
Metadata appears in several places inside a record. The paths below are
relative to the record object, not to an API response wrapper such as record
or result. Business fields remain alongside these members.
Top-level members
| Path | Type | Meaning |
|---|---|---|
_id | string | Record identifier |
_source | array of objects | Recorded provenance; each entry has _src and optionally _src_id |
_metadata | object | Stored processing and record-state information listed below |
_search | object | Evidence attached to a search result |
qualityState | string | Derived measurement state: CURRENT, STALE, or UNCALCULATED |
_source[]._src is the source name; _source[]._src_id is its optional record
reference. In Aurelia, the CRM load records sample-source-crm and CRM-0001
for Ana. The source and reference are strings. The dataset can constrain their
vocabulary and validation through its provenance definition. Golden uses the
reserved source _golden for records created in Golden.
No API response carries _audit, including an expanded record response, and
a write request carrying it is refused. Read a record’s audit events through
the dedicated record audit endpoint.
Only a destination receives _audit, inside its payload; see the
product event reference.
Members inside _metadata
Each key in this table is nested: for example, the quality score is
_metadata._quality, not a top-level _quality field. Members can be absent
according to configuration, processing history, and measurement state.
| Key | Type | Meaning |
|---|---|---|
_errors | array | Legacy processing errors; omitted from responses with current quality measurements |
_merged | array of strings | The records this record absorbed in merges. It has this one meaning everywhere it appears: on a live record, on a history row and in a product event. A merge survivor lists the records it absorbed — never its own identifier, and never an identifier it continued. A record a merge absorbed keeps its own list, and does not point at its survivor: read _merged_into for that. A record that a merge absorbed before Golden 3.2.0 carries no _merged in a history read: the list it had was not kept. Absent, never null, when the record absorbed nothing |
_merged_into | string | Only on a record a merge absorbed: the survivor’s identifier. A live survivor, a deleted record and the survivor of an undone merge never carry it; a survivor that a later merge absorbs does, because it is then an absorbed record. Read _merged_into, not _merged, to follow an absorbed record forward |
_merge_id | string | The identifier of the merge that produced this record. Present on a merge survivor only; absent on a record no merge produced and on a survivor produced before Golden 3.2.0. A record that a later merge absorbs keeps the identifier of the merge that produced it, not of the one that absorbed it. A destination transformation drops it |
_unrelated | array of strings | Identifiers recorded as unrelated or disconnected. An undo copies the separations recorded against the survivor onto every record it restores, on purpose: a judgment recorded while the survivor existed is kept rather than lost |
_operation_bucket | string | The candidate-group operation of the last write: merge, remove, disconnect, or unmerge. split appears only on records and audit events written before Golden 3.2.0, which no longer splits groups. Absent on a plain insert or update. A deleted record carries remove in the delete’s API response and in its history row as an API read from history shows it; its destination event carries no _operation_bucket |
_quality | integer | Record-quality score from 0 to 100, when measurable |
_quality_facts | array of objects | Findings for a current measurement |
_quality_error_count | integer | Findings of severity ERROR for a current measurement |
_quality_warning_count | integer | Findings of severity WARNING for a current measurement |
_quality_available_points | integer | Weighted total against which the score was calculated |
_quality_definition | string | Definition against which quality was measured |
_quality_calculated_at | string | ISO 8601 time of the measurement |
On a live record, _operation and _operation_bucket describe the last write only: a plain
update to a merge survivor clears _operation_bucket, and the survivor still lists what it absorbed
in _merged.
_unrelated, _merged and _merged_into name identities, not stored rows. On a table without
history, an identifier in _merged can name a record that a later load re-created under the same
identity.
Which record survives a merge
A merge’s survivor takes the identifier the dataset’s identity rule computes from the merged values:
| Dataset identity | Survivor |
|---|---|
DEFAULT | A new record, under an identifier assigned when the merge runs |
COLUMN or SCRIPT, the merged values compute the identity of one of the merged records | That record continues: it keeps its identifier and is updated in place |
COLUMN or SCRIPT, the identity of a record in history that leads to one of the merged records | That identity comes back as the survivor |
COLUMN or SCRIPT, an identity nothing holds | A new record under that identity |
COLUMN or SCRIPT, an identity held by a record outside the merge | The merge is refused with 409 and existingIds; nothing is written |
COLUMN or SCRIPT, no usable identity | The merge is refused with 400 |
On a COLUMN or SCRIPT dataset whose merged records have different identities, which one survives
is the merger’s choice, through the values it produces; under a RANDOM merger it can vary from run
to run. The merge preview and the duplicate group publish the identity the merge would take as
mergeIdentity, and the merge screens disable a merge that would be refused.
A new survivor is fed by no source
A survivor that is a new record — always on DEFAULT, and on COLUMN or SCRIPT when its
identity is new — is fed by no source. The merged records’ identities are now in history, so every
later load skips their rows, and nothing updates the survivor except a person or a system that sends
the survivor’s own identifier. This is the model, not a gap: the new record is none of the originals
any more.
A survivor that continues one of the merged records keeps being fed by that record’s source.
On a table without history, nothing records where the merged records went: loads do not skip their rows, and a later load brings them back beside the survivor.
The merge confirmation says which of these applies before the merge is made.
What _updated means
| Where you read it | What it is |
|---|---|
| A live record | The time of its last write |
| A history version | The time the version was archived — by the delete, the merge or the undo that wrote it — not the record’s last write |
| A row a Golden-table sink wrote | The time of the sink’s last write to it, like any live record. The event it applied last is in _sink_event |
A remove event | The time of the operation that removed the record. An insert or update event carries the record’s own _updated |
Quality metadata
_quality is 0 to 100, not 0.0 to 1.0. A record scoring 78 reports
78.
Read the score with the top-level qualityState:
| State | Meaning |
|---|---|
CURRENT | The stored definition matches the table’s current quality definition |
STALE | A measurement remains but its definition differs from the current one or is missing |
UNCALCULATED | Neither a quality definition nor a measurement has been recorded |
This state is derived when reading the record. A comparison between
_quality_calculated_at and _updated does not determine it. A dataset change
can make a stored measurement stale without changing the business values.
When the state is not CURRENT, the response omits _quality_facts,
_quality_error_count, and _quality_warning_count. It can retain the score
and its available points alongside the state. Missing findings are not a
successful validation, and an absent score is not zero. CURRENT alone does
not guarantee a numeric score: a record with no measurable values can have a
current measurement without one.
Each entry in _quality_facts carries:
| Field | Meaning |
|---|---|
severity | ERROR or WARNING |
messageKey | Stable identifier of the finding, for example quality.validation.email.invalid |
messageArgs | Values interpolated into the localized message |
schemaPath | The dataset column the finding is about |
recordPath | Where in this record it occurred, which differs from schemaPath inside nested or repeated structures |
Key on messageKey, never on rendered message text. The text is localized
and can change between releases; the key is the machine-readable contract.
How _operation and _operation_bucket combine for each kind of change is
listed in the product event reference.
Search metadata
_search can contain index, match, type, rank, exact, origin, and
sourceId. It describes one hit per result, the record’s best, and is
derived from the record’s own key at search time.
match is the normalized value Golden indexed, not the term submitted.
Treat rank as an ordering signal rather than reproducing Golden’s ranking
behavior, and read origin before claiming a record matched a value. The
values of type and origin, and when exact is true, are described in
Read the search evidence.
Input and optional fields
Do not write processing results as business data. qualityState, _metadata,
and _search are not client-maintained business fields. _id and _source
have their own input rules, described in the
record write reference: build a write request from them
rather than copying a complete read response. Clients should
tolerate additional metadata properties so compatible releases can add
customer-visible context.