Canonicalization

Canonicalization is the process of deciding whether a person, organization, or location found in an article matches a Stylebook record, is new, or needs an editor's judgment.

This happens when an Agate flow saves results through Backfield Output. Each extracted entity has one of three outcomes:

Outcome	What it means
Link	Attach the article evidence to an existing canonical record
Create	No good match exists, so a new canonical record is created
Review	The match is uncertain, so the item enters a candidate queue

Linking to an existing canonical does not overwrite its editorial fields with the latest article extraction. The article-specific record and evidence are attached while the canonical remains the newsroom's authoritative identity.

How matching works

Matching rules differ by entity type:

• People use normalized names, aliases, and affiliation signals. Similar names are not enough when evidence suggests different people.
• Organizations use names, aliases, acronyms, and organization types. Ambiguous acronym matches are held for review.
• Locations use names, location types, addresses, jurisdictions, and available geography.

Inactive records are not linked automatically. Conflicting identity information can prevent a proposed match.

Rules and AI suggestions

You can run canonicalization in two modes:

• Rules uses fixed matching logic.
• AI-assisted asks a model for help when the rules are unsure.

AI assistance cannot bypass the identity safeguards. Uncertain or conflicting results still go to an editor.

Candidate queues

Unresolved people, organizations, and locations appear in separate candidate queues. A queue can include work from several projects using the Stylebook and shows which project produced each candidate.

For each candidate, an editor can:

• link it to an existing canonical;
• create a new canonical from the article record;
• defer it for later;
• inspect its evidence, suggested matches, and similar records.

Deferring a candidate moves it out of the active queue. It does not delete the article evidence.

AI review can suggest link, create, or defer actions. Nothing changes until an editor accepts a suggestion.

Canonical cleanup

Canonicalization handles incoming article entities. Stylebook Review checks records already in the catalog. It can flag duplicates, questionable links, and location geography concerns. Editors can merge records, keep them separate, delete an empty record, or dismiss the issue.

Merging moves the linked article records from the duplicate into the record you choose to keep, then deletes the duplicate. Review both records before confirming; Stylebook does not provide an undo action. The Recent view can show that the merge occurred, but it cannot restore the deleted record.