One row per curated similarity edge (grain match_id) between an opportunity and a precedent job reference, with similarity_score (0–100), a prose relevance_note explaining why the precedent matters (method reuse, same client, overlapping test scope), and referenced_in_rfq marking edges the client's own RFQ cited. matched_proposal_id joins the precedent to its proposal so line-item pricing can be re-aggregated.
- mediumTreating similarity_score as a model output
Scores are curated/heuristic (estimator judgement or synthetic generation), not embeddings similarity. Use them to rank, not as calibrated probabilities.
- mediumAssuming every match resolves to a proposal
matched_proposal_id is NULL when the precedent reference has no proposal row (e.g. a method-development reference cited in RFQ prose). ROM reconstruction is only possible for non-NULL matches.
- “Which past jobs are comparable to this opportunity, strongest first?”
- “What ROM did the closest comparable carry?”