Delta ingestion

The model is maintained, not regenerated. New information becomes proposed changes against approved, never a fresh model -- stable ids and the staging overlay exist precisely so a re-run produces a reviewable diff.

Pipeline

1. Intake first. New documents go through ea-intake like any source: chunks, facts with verbatim quotes, entity resolution against the existing facts/entities.yaml (this is where "the new WMS" gets recognised as wms rather than becoming a twin), validate-facts, coverage. A PDF, Word, PowerPoint or Excel file is converted with easkills intake-doc first, never by hand.

   Check the open questions against the new material. An answer often arrives unannounced, in the next interview or export. When a new fact answers an open entry in facts/questions/, close it: status: answered, the date, and answeredBy naming the fact. A question answered and left open keeps appearing in kpi and in the architecture description as an unknown the repository in fact knows.

2. Ask the tooling what is new:

easkills delta --root <repo>

Two mechanical lists: entities with no model counterpart (candidate additions) and facts no model concept cites (unused evidence). They are candidates, not defects -- your judgement decides which are in scope.

3. Propose the delta in staging. For each accepted candidate: new elements and relationships in model/staging/, citing facts (provenance: [{fact: ...}]). For changed reality, re-propose the existing id with updated content -- the overlay treats it as an update. For things the sources say are gone (decommissioned, replaced): propose the retirement explicitly -- set properties.lifecycle: end-of-life (or remove the element in the staged copy of its file) and say so in the report; silent disappearance is how models and reality drift apart.
4. Validate and hand over. validate --zone staging, three-repair cap, then report. Promotion stays human (ea-approve).

Contradictions are findings, not merge conflicts

When a new source contradicts approved content ("the portal now holds stock levels" vs the approved claim it does not), do not quietly pick a winner. Propose the update in staging citing the new fact, and flag the contradiction explicitly in your report with both quotes -- the human promoting decides which reality is current. If the old claim came from an older document, that is expected drift; if both sources are current, that is a clarification question for the owners.

Reporting back

Lead with the diff shape: N additions, M updates, K proposed retirements, and the contradictions needing a ruling. List what you deliberately did not model and why. The unused-facts list that remains is fine -- not every fact becomes architecture.