Brief
Author a citation-required eval set over the RAG corpus and score your m1 retriever + generator with it. Report groundedness, refusal accuracy, and cost/latency per item class.
Dataset
Golden benchmark set (starter): grounded factoids + OOD/refusal items + conflict/stale traps. Train distribution differs from the graded split — report both, tune on neither blindly.
Starter guardrails (plan 88)
Use the judge harness with pinned judge model + temperature; log position-shuffle; no hand-editing scores after the run — rerun instead.
100-pt rubric
Eval set quality (coverage x difficulty) 30; gate+scale scoring discipline 30; cost/latency report 20; conflict/stale handling 20.
Deterministic checks -> LLM review
Checks: all items scored; gates computed before scales; cost columns present; judge config pinned. LLM review grades anchor calibration and bias guards.