Brief
Defend groundedness and spend: eval report (gates + scales + bias guards) and a cost dashboard (per-item tokens/USD/latency, outlier sign-offs) rendered from the JSONL logs — no hand-edited numbers.
Dataset
All suite + eval logs from m0-l4 through m3-l16, joined on run/item keys.
Starter guardrails (plan 88)
Dashboard reads the logs, never pasted figures; sign-off entries for any 10x-cost trade; judge configs pinned in the report appendix.
100-pt rubric
Groundedness evidence 30; cost/latency completeness 30; bias-guard proof 20; defense clarity (decisions trace to tables) 20.
Deterministic checks -> LLM review
Checks: dashboard renders from logs (spot-check 3 rows); sign-offs present for outliers; judge configs pinned. LLM review grades the defense narrative.