Evaluate
Retrieval quality and answer quality are separate failure modes, so they are scored separately. The retrieval pass costs nothing and spends no tokens; the architecture pass spends real ones.
Gold set0 cases
Retrieval
recall@k, precision@k, MRR and nDCG for each retrieval mode. nDCG is the one to watch: it penalises burying a relevant passage at rank 6 where context truncation may drop it.
Architectures
End-to-end: route accuracy, substring answer assertions, citation recall, cost and latency per arm. The no-tools baseline is included on purpose — without it you cannot tell how much of a score came from retrieval.
Cases
| # | Question | Route | Relevant refs | Must contain |
|---|