Turns production AI-agent traces, logs, CSV/JSON exports, or incident notes into an evidence-linked observability and FinOps brief. It reconstructs tool-call chains, separates measured facts from derived values and hypotheses, and produces a formula-level cost ledger with confidence labels and ranges instead of false precision. The output covers retry waste, duplicate calls, failed loops, context and retrieval issues, long-tail latency, workflow segmentation, and reliability findings. It then proposes small, rollback-ready experiments with guardrails, validation metrics, and stop conditions, plus a reusable instrumentation schema. It explicitly asks for missing prices and context, quotes source evidence for material claims, preserves unknowns, and includes a cannot-conclude section; it does not pretend to access systems or invent rates, owners, baselines, or business outcomes.