Sub of Epic #318. documents/audits/AUDIT_RAG_LAYER.md Phase 3 (size M).
Scope: RAG eval inside the shared src/strategy/eval/ package (#205), not a parallel one.
- Wire a reproducible RAG retrieval eval: ~30 queries including the 3 production question shapes, computing P@k / MRR + a wrong-year rate + a citation-match rate. Revives the one-shot unwired N30B notebook (ML-eval E-11).
Deliverable: f1-eval rag (via the shared harness) reports retrieval quality + wrong-year + citation-match over a fixed query set.
Sub of Epic #318.
documents/audits/AUDIT_RAG_LAYER.mdPhase 3 (size M).Scope: RAG eval inside the shared
src/strategy/eval/package (#205), not a parallel one.Deliverable:
f1-eval rag(via the shared harness) reports retrieval quality + wrong-year + citation-match over a fixed query set.