Identities and mentions

LPGCast computes every identity before it builds an edge, the same way graflo does for any other source. How depends on how the vertex type is identified:

Identity of the type	Where its key comes from
natural (identity: [email])	the text: the model must give the property
hash (hash_identity_properties)	graflo's digest of those properties
assigned (assigned: true)	a UUID derived from the schema and its version, the episode's source_id, the type and the model's ref
blank (blank: true)	derived the same way

The derived UUIDs make a replay reproduce the same graph. A fresh model call that names its refs differently derives new ones, so assigned and blank vertices are not idempotent across re-casts; natural and hash vertices are.

Mentions

A text often refers to something it says nothing new about: "the customer with email jo@example.com raised HD-4415". The model marks such an entity as a mention. A mention is never written as a vertex:

• known by its primary identity, it is simply an edge endpoint;
• known only by a secondary identity (secondary_identities in the schema), the edge waits in pending.jsonl, and its endpoint is looked up in the database at write time. A lookup that finds nothing, or more than one vertex, leaves the edge out and is counted in write.json.

A mention with neither is an identity_insufficient finding. That includes any mention of an assigned or blank type ("the work order from last week"): there is nothing to look it up by. Such an entity is best left out, since writing it would mint a duplicate; the benchmark's mention family measures how often a model does.

Edges

An edge is identified by its endpoints and its relation. Give it identities in the schema (identities: [[source, target, since]]) only when several edges between the same two vertices must stay apart; its other properties are written, never matched on.