Which retrieval path is failing?
Evaluate route choice, traceable evidence, and completion separately. Trust requires all three.
Labeled trace CSV
Six representative samples are ready.
A correct answer can still take the wrong path.
Compare the labeled route, selected route, evidence count, and final status before trusting aggregate quality.
Canonical evaluation
Route and evidence audit
Result is useful before export.
0samples
0route correct
0evidence
0trusted
0median ms
Pred RAG
Pred SQL
Gold RAG
0
0
Gold SQL
0
0
Trustworthy pass
A sample passes only when the selected route matches the label, at least one evidence reference is retained, and the final answer status is answered.
Evidence boundary
This tool evaluates imported labels and counts. It does not run RAG, SQL, a model, citations, a database, or live traffic, and it cannot prove the supplied gold labels are correct.