DR-148: Entity-mention ownership follows the retention class
DR-148 — Entity-mention ownership follows the retention class
Section titled “DR-148 — Entity-mention ownership follows the retention class”Context
Section titled “Context”The corpus reframe (R1) gives every source binding a retention class, and its rule for ingest-once sources is explicit: the pipeline engine assumes “source disappeared → tidy up its rows”, and for ingest-once sources that is exactly wrong — the extracted knowledge must outlive the source. The id-434 reshape makes mention rows engine-owned declarative target states, which is precisely the ownership that would reconcile them away when an ingest-once source (a meeting transcript, an onboarding Q&A document) leaves the walked tree. id-138 already built the per-class what-survives-the-engine contract for records ({138.11} one-shot ingest-once extraction path, {138.16} promotion-boundary regression); mentions were not demonstrably inside it. Ruled on the S554 review board (owner pick: full per-retention-class design, not walked-only).
Decision
Section titled “Decision”Mention ownership is decided per retention class of the source binding, not globally:
- keep-and-watch / live-connected — engine-owned declarative rows: the id-434 shape; the source persists, so engine reconciliation is correct behaviour.
- ingest-once — mentions cross the promotion boundary: extraction lands engine-owned, a promote step moves the rows out of engine ownership before the source retires, and “walks never touch promoted records” (DR-026/id-138) holds for mention rows exactly as it holds for q_a_pairs.
- external-referenced — no walk, no mentions from bytes; the reference_items lifecycle (DR-124/DR-132) is the whole of it.
Alternatives considered
Section titled “Alternatives considered”- Engine ownership for everything. Loses ingest-once knowledge on source retirement by design — the exact failure the corpus reframe names.
- Promotion for everything. Wrong for walked sources: DR-140’s point is that the engine’s tracking record and the database agree; promoting walked mentions out of engine ownership re-creates the divergence deliberately.
Consequences
Section titled “Consequences”- id-450 executes the verify-and-extend: establish whether mention rows sit inside id-138’s contract today (presumed not — they are engine-declared) and extend {138.11}/{138.16} to them. id-138 is archived and is not reopened.
- {434.2} designs the walked shape and names the ingest-once seam so the phase-2 component does not foreclose it; today’s corpus is all keep-and-watch, so id-434 is not blocked on id-450.
- id-420’s publication facet is the surface where mention promotion would appear to a reviewer; coordinate there rather than minting a new surface.